DeepSeek v4 Pro did same for $1.7 in 7 responses (peak-off times).
For doing much complex task, I would be super nervous about using Anthropics models.
For AI agents, we have been mostly using GPT 5.6 Terra.
So far not looking back to Anthropic. They lost me.
and resources or tech stack tips from HN?
I've been running Qwen27B Q6 with 130k context at 30tps. It's not bad.
I think with an "autopilot" mode in pi, and sub agents to handle context window management better, I'd envision you could get close to unattended workflows. I haven't quite gotten that far though.
Negative is you have to do it all, and the temptation to tinker is real.
Did they not learn? Performance is a feature
I can respect the guardrails - I also can see why OpenAI may not have much control over their models - but I need an AI who will do whatever I ask and not play judge and jury.
I mean, who says "screw you" to requests to get 35+ year old vintage computers working? Claude, that is who. Its guard rails are so stupid. I hear people trying to do simple mailing list management hit it too.
I am just about done with them.
At this point, Fable is really just an experimental model (as it should be). It can do very useful things, but is it a broadly good general purpose model? Definitely no. Most users don't have problems hard enough for Fable outside of coding very large and complicated projects. Most users don't have 45 minutes to accomplish a task Sonnet can do well enough in five. There's not a PowerPoint in the world where Fable is the right tool to build it.
The secret sauce is going to be in letting Sonnet decide to delegate to Opus and Fable when they're the right tools for the job. But you can't train a model to do that until the bigger/better models exist and you can observe how your users actually take advantage of them. I'd bet money that's exactly what the rlhf going on at Anthropic looks like right now.
Economically, it makes sense. Sure, on paper you want users burning as many tokens as you can. But pushing users into burning tokens and taking a long time and getting a meh result is far worse then giving them the "fast and good enough" solution that occasionally burns more tokens automatically when the problem requires it, and getting a higher quality result out when you do. From an infrastructure capacity perspective, this is the dream: you stop measuring cost [for Anthropic] per token and measure cost per outcome, allowing you to use less hardware to accomplish the same abstract units of work.
It's unclear if Mythos2 or 3 or whatever they're calling their next model will be an improvement for most common enterprise use cases.
LLMs can't solve basic things (writing non-slop documents, understanding context without massive handholding) and for coding other models are quickly becoming 'good enough' without the same cost and nannying.
That's why Anthropic is 'stealing' workflows.
But it turns out it's much harder to push adoption when your users don't really want to use your product.
Code was a unique use case where the code luddites were loud but a minority - most people don't want to update 300 cases of variables across their code base for a name change. Most don't want to write unit tests.
There are a few use cases where that will happen (law is next, maybe quant finance) - but otherwise most companies are throwing money into a pit and getting 0 return.
It's a very interesting race and state of affairs, but Kimi K3 and likely the next DeepSeek models will put the high price token affair to rest.
Unless of course, mythos / next model really does solve some universally applicable problem that people want it it to do.
I've now bought the $200/month plan from OpenAI and simply ran `/status` in Claude to get my session ID, then asked Codex in the same directory to "Please take over the work started by Claude Code with session ID `[…]`." This seems to work like a charm. It also seems that the Codex weekly and monthly budgets are more generous?
Unless Anthropic adjusts their customer-abusing behavior I think I will end my subscription with them soon.
Most of the people pushing this are just hoping that they can cash out before the hype pops and financial gravity crashes the party. Sam Altman recently claiming that the singularity is here is so stupid on its face he should just be treated as what he is, a huckster.
None of this stuff ever made any sense on what it was being sold initially. It was always insulting that the media and business leaders tried to argue that the tech could replace entire call centers or vast swaths of entire industries.
People keep arguing, but it will or it has based on extrapolating certain, reasonable use cases. Klarna has shut up about replacing call centers with bots because Markov chains with memory only can do so much.
I’m happy with Claude. If they become (bigger) jerks, I’ll switch to something else. I don’t ha e the energy to praise Anthropic today and I won’t have the energy to demonize them tomorrow. The emotional investment people have for/against these companies feels like celebrating or being offended by the weather.
TBH I don't think any of that is unique to the IT industry's relationship with the Valley. Other technology-driven industries have a similar worship-ish relationship with a few rarified businesses. But the culture of the IPO exit accelerates all the most short-term motivations to do anything.
The whole "we're going to replace all of your workers with an agent while also making AGI happen" thing was a bad idea in the first place; one that could not have gotten anywhere outside of Silicon Valley. People just want tools that they can deploy at scale, not to be your beta testers for the singularity.
outside of coding scenarios, anthropic is just not quite the best choice for regular folks.
depending on the cloud/office stack used by the company, the userbase is better poised to use the native ai offering than moving to anthropic and connecting everything with them. same goes for third party guys trying to tie it all together.
not to discount them or their features, but beyond small companies or startups, we don't see the same levels of adoptions now nor maybe in the near future.
those asking us to build genai solutions for them also do not choose anthropic first over the main model offered by their current CSP. anthropic made itself available across the board finally, but it was already too late for most projects planned out for the upcoming quarters. also, we don't use the absolute biggest model nor instantly migrate to the latest one. the insane price jumps do not help with the same. especially when we quote something that may no longer be true down the line.
aside: good luck to the FDE folks hired by the frontier labs. i wonder if they would really get to handle large company-wide projects in practice.
I become more disillusioned every day.
It seems like, by having the models write Python code, they tend to write Python code like an average developer. Which is to say, quite bad.
Add in the complete failure of the models to adhere to instructions in Claude.md, memory files, and added multiple times in prompts, I’m wasting huge amounts of time fixing bad design decisions that the model just slips in.
No shit we're all paying 清冲 Flash to do the grunt work. Turns out though, paying 清冲 Flash a few more cents does exactly what Ivy Wasp Pro Mythical does. Crazy how that works.
- Fable, nerfed or whatever, too expensive and not fully included in subscriptions.
- Opus, neurotic (excessive) slop machine.
- Sonnet, way too token hungry, cost much more than 4 series.
Their almost daily outages does improve the experience. Anthropic really have messed up this year.
Bro if you want us to spend more on fable let us use our whole damn rate limit for it. Enterprise is a different ballgame obviously but for the subsidized Claude code users they literally cap it
Anthropic is the fastest-growing software company in history. They have no issues "attracting users"
This wouldn't be so bad if it didn't make mistakes periodically, especially when it's about to tap out. I get the sense that if I upgraded to the $200 subscription it would get me a lot more usage, but it would still run into these issues anytime I sat down to work for a few hours.
I'm just using medium effort, so it's not like I'm on high all the time.
Saying "people are not buying that many Rolls Royce, instead they buy a lot of normal cars" is kind of duh. Anthropic is clearly still operating at inference capacity.
At this point, every model writes code about as good as I'll need for the stuff that I'm working on, and with the right harness, I'm perfectly happy to let them cook, until it comes up with something working. Whether it takes 10 minutes or an hour really makes no difference to me.
If I need to write all the specification so it follows it, I might just write the code or use a cheaper model.
Opus 5 and Fable 5 have been quite disappointing.
I used sol, terra and Luna and I think they and the first two are good for first code reviews. Luna is not much better than deepseek V4 flash
This is just a case of competition working: Fable (and Opus 5.0) just aren't as good as Anthropic believes, and there's no switching cost, so everyone's moving away.
if people dont see why they need a model this smart, they probably arent using ai enough
I am not doing awfully complex tasks though. I imagine a lot of other people are in a similar boat, either switching from Claude to ChatGPT or even just min-maxing DeepSeek V4 Flash 0731 or similar.
Then on top of that opus slurps tokens like Anthropic is afraid they’re losing money. Read 200 lines of file, +20k tokens. Excuse me?
Anthropic in particular is much more compute-constrained than OpenAI and SpaceXAI and has relied on partnerships to provide inference. This reality factors into their pricing and usage limits (they started 'adjusting' the 5-hour limits during peak hours, and it certainly wasn't an upward adjustment). Accordingly, this is presumably what Anthropic wants, given they develop and release the lower-end models, suggest users use them in various nudges within their product, position the bigger/more expensive models as "For the most complex tasks" in their UIs, and so on.
I think the problem (for Ant/OAI) is that there is no sensible lockin or moat. LLMs are essentially interchangeable and stuff like a harness doesn't offer enough value on its own for someone to be locked into using one of them.
Now with the onslaught of the Chinese models that offer almost the same quality for much less money they have a very serious problem on how to proceed. Investors now might be looking through rose tinted glasses but their patience has its limits.
Enterprise is a whole different ballgame IMO with countless technical, compliance, and employee adoption considerations that add friction to switching. It's also where both Anthropic and (as of last week) OpenAI get the bulk of their revenue, and, incidentally, the venue where US Government regulations on Chinese models would have the most impact.
This is the first time I've felt, and I use the word *felt* since I don't have a suite of benchmarks or any sort of material approach towards comparing models, that Opus has declined in quality compared to before. Primarily I think its powers of deduction and understanding, even on xhigh, have become much worse. Before, being vague and providing a simple prompt would be enough, it could deduce and expand the details it needed, plus ask you clarifying questions, now this is no longer the case. A concrete, personal example, for a personal project, I've asked it to setup ssl over local IP. I didn't go into too much detail in the prompt as there are many approaches it could take and I didn't care too much to choose. It did horrible. The first thing it did was say the best lightweight approach is to add a reverse proxy. I'm like ok, makes sense. Then after asking it to proceed, it went and added a bunch of config to my golang service and didn't even setup a reverse proxy even when it said that is the way to go. It even said it didn't set it up lol. Then after I told it to do so it failed building the config in a way it was asked of it (support LAN IP and tailscale IP). Etc etc...
When Fable came out it was huge, the benchmarks told the story, and the story mostly matched the experience. It felt, again, intentionally saying felt, like it was miles ahead. Now benchmarks say that there are many models that are close, but in actual use Fable still *feels* much better. I think benchmaxxing the new open weights models is ruining the value of benchmarks, if they ever had any. When you actually put them to the test you see 500k tokens of reasoning with "Actually..." and "Wait..." in every third paragraph of their reasoning trace.
The price for Fable is definitely too much for any personal use now that it's no longer included in the subscription, and GLM 5.2, Deepseek Flash and Qwen 3.8 served locally or via cloud provide a lot, requiring a bit more babysitting though. Considering the price of Fable, my 5k USD Epyc server would pay itself off in less than a year if I used Fable or Opus in the same manner so at least for me the decision seems easy. And considering the point I'm poorly trying to make, that Opus doesn't feel like frontier anymore, this is probably the last month of my Claude subscription.
Fable is not used because it has extra retention requirement that corps can't sign off so it stays disabled for everybody in many cases.
It's also not winning on day to day work against Opus 5, which is simply available as there are no extra retention requirements and no extra paper work to do with legal.
Devs also don't like that Fable refuses to work on half of their prompts.
Then comes better pricing and on-par capabilities from competition.
In what universe? Hell, Opus 5 seems to be worse than Opus 4.6 at almost everything.
Computing technologies are relentlessly deflationary. If the value of their TAM wasn't a fraction of a manual process they replace, they wouldn't have a productivity advantage. And the amount of TAM per unit often declines over the life of that product category.
I would be unsurprised to find investment in data centers to be 10X what was really needed. The same goes for where the LLM S curve starts to flatten.
Nobody has complained and seems like for every use case we have Opus is more than powerful enough, especially with Opus 5
I find it funny how OpenAI got caught lacking for a very brief window, but it turned out to be a very critical turning point.
Like a guy that that's at the top of their game the entire year, and the one day they have the flu, the CEO does a surprise performance review.
Sure there are Microslop and Oracle db users but most of the world we live in is Postgres and Linux. That's why I think most companies will run llm's like that.
1. They acted as if they were so far ahead capability wise that they could stop listening to their users.
That is basically it. It is an extremely common belief that Opus 5 acts like a condescending wannabe-thought-leader, yet Anthropic's response to this is largely been "You're using it wrong. Try deleting all your config files".
They released a "concise" output format, but pretty much swept it under the rug, despite it being one of the loudest complaints about their models. It also, generally, does not work as described.
Not only did Opus 5 get much more difficult to work with, but it got, from a customer view, significantly slower. The token rate might be the same, but if every interaction takes 50% more tokens, it's 50% slower.
All this time OpenAI has released a slew of new models, lowered the price on them (which, bluntly, 80% of user count don't actually care about since they're on subscriptions), increased subscription capacity, and increased response speed.
It's not about cost. It's about Anthropic being the frontier-lab version of the marathon runner who decides to celebrate to early, and then loses the race.
If you are disproving Jacobian Conjecture it makes sense to be on SOTA, but for writing Golang and Typescript, faster sol/fable/opus class models are imo more likely to get user interest than the latest frontier.
The frontier models may be better, but who cares if the last generation of models are plenty good enough for what you want to actually do?
And the best local models are quite competitive with those older lab models.
Then all the mythos warnings happened, then the fable access issues started.
At that point, I just wanted to try something else because I didn't want to feel vendor locked by anthropic.
Then I tried codex on the $100 plan. I've never been happier. It does everything claude can do, without any token limit issues or anything for me.
I don't see myself going back to claude anytime soon. I also highly recommend folks to not vendor lock yourself to a model. You need to try multiple models for yourself and see what's possible. There should be zero loyalty to any frontier lab because of how fast things change and how volatile model access, uptime and customer experience can be with these things.
You can't make it an indispensable dev tool if devs cannot rely on it. They will absolutely jump the service to s more reliable one and that was Codex for me.
P.S. Claude is indeed down right now for me.
Also Sonnet 5 was released June 30th, seems to be grouped in with 'other'?
https://ramp.com/data/ai-index
(click on model market share)
I have tried fable, gone back to sonnet, tried to use agent mode to mix and match and various other combinations. If I decide that Fable is the only one that can save me hours or days of time, I will go back to it and pay the cost. If the cheaper or free ones do well enough, I will stick with them.
When you see the price of hardware needed for these models, we are generally not paying very much towards that price so I expect prices to start ramping up as the honeymoon ends and these companies' investors get itchy for their ROI.
I've wondered if this is part of why we don't see the reasoning traces for Anthropic's models before -- Open models might just be accurately surfacing how the sausage is made.
> good at its job just use it instead of re-inventing the wheel
Exactly why is everyone reinventing a harness every month. There will be Microslop / Oracle harnesses and there will be 1 or 2 open source ones that win.
They have tried to find the highest that the market pays for sota models; however, on the consumer side, this is just too confusing and unsettling:
"You can only use Fable for a week as a part of your plan" "Be ready! You have to start paying per token!" "Nevermind! we extended it for a couple more weeks" "Wait, now it's up to half your usage" "Ok, now its..."
Most people want to not care. We want our AI like electricity -- Kind of just there no matter how easy/hard is for the supply. You don't want your electricity company to be on the brink of cutting you off any second.
That's Anthropic. You don't feel they want to give you a dependable service for an, albeit premium, price. It's a constant bargaining game. That forces people to look beyond the walled garden. There, they find models that are fine... and without the shenanigans.
It also didn't help that the government yanked it which adds another source of anxiety since OpenAI is on much better terms with the administration and the administration seems corrupt enough that they would mess with Anthropic if they got a big enough donation from OpenAI.
But anyway after Sol entered the picture, I don't think Anthropic can get away with this as much and I also think they're going to face a massive backlash from Max subscribers if they do end up ending the +50% promotion at the end of the month because Sol is a Fable peer and priced very competitively.
Then she hired a VA in the Philippines. Anthropic promptly banned her account without warning once the VA connected to the account. It took her weeks to get her account reinstated, at which point she had already moved on to OpenAI.
The thing I miss most about programming is flow, and the constant bouncing between terminal tabs sucks. I’d love to do one thing at a time, with Fable, quickly.
Which tells you everything you need to know about OpenAI.
Every time I get "you used your quota, come back in 3 hours, or 2 days" -> that is experimentation time with their competition, leading to changed service plans. When they said "claude -p" will be billed at API pricing even for plan users I moved my harness off claude. After I integrated codex, then it was never going to be a full claude project again.
What business encourages users to try their competition and adapt their usage to the competing products?
I've never run a local LLM model before. Certainly won't take as long to iterate on this.
If you are selling something, and losing $10 on each sale, you also would want to limit how much you sell.
I mean, sure, you are losing money on each sale so you can landgrab, but you still have to balance the land-grabbing with how much money you can actually lose.
I guess this is why they're pushing Claude code hard (not supporting agents.md, not allowing third party harnesses, etc) but when switching to another provider is as easy as opening a new terminal and typing omp/pi/codex your moat is effectively zero.
They can compete on price, quality or value but anything else is just madness. Currently they (arguably) own quality but this won't last.
They are high on their own supply. The people running these companies are delusional imbeciles who have been placed in charge of billions of dollars.
The breaking point for me was the privacy violation. They've been fingerprinting every request and violating users' privacy hoping no one would notice. Too bad, someone found out and that was the day when I cancelled my subscription.
I'm strongly considering biting the bullet and just ditching my $200/month Claude Code plan for the Codex one instead, especially because I keep running into my weekly limits (even sticking to Opus.)
1. No 5 hour usage limit
2. Weekly usage gets reset CONSTANTLY. It's crazy. The longest I've ever seen it go without a reset is maybe 5 days?
3. I don't feel like OpenAI is constantly trying to fuck with me. Unlike Anthropic. I would way rather have Sol all day every data, consistently, than a slightly better Fable for like, 1 prompt every 5 hours, and only when Anthropic decides to not treat me like a cyber criminal. Believe in yourself as much as Claude believes your CRUD app is going to hack the pentagon.
4. Getting access to image generation, though I don't use it too much, is a nice perk compared to Anthropic.
edit: Should mention that I had like 4 banked manual resets as well. It feels like OpenAI wants me to use their product, whereas Anthropic wants my money while giving me a nerfed experience
I’m sticking with gemini for now given the 20TB cloud storage and youtube premium that comes with it but I’m still open to switching to codex (which I’ve had a long term plus plan for). If google keeps delaying the pro models for much longer or makes them excessively expensive, I’m likely to switch out.
The consumer side cheap monthly plans exist for the same reason companies like Cloudflare and Vercel have a free tier: When it’s cheap and easy to get developers familiar with the tools, they will push their companies to pay the real money for those tools.
It’s a hard balance with LLM serving because you can’t really make it free. $20/month is close to free, but the $200/month plans are in a difficult place where they’re big enough that many small companies pay for $200/month plans for their employees and ignore the enterprise features you get with the full expensive arrangements. So the companies are continually adjusting the $20-$200 plans to keep them from being reliable options for businesses, which is where the real money is.
There’s a short sighted cheering on of the 3rd tier and lower companies offering lower rates, but we’re already seeing them ratchet up the pricing and keep larger models closed after they get market attention.
Sorry, but people are cheering on Chinese companies (of whom your are unduly dismissive with your '3rd rate' comment given how good GLM-5.3, Kimi K3 are) not only because they are more economical, but also because they do not constantly refuse to do legitimate tasks and provide you with the weights for self hosting these models.
The period was also marked with many billing bugs, like spending people’s usage credits for included Fable for a few hours (gave me a huge shock), but to their credit they refunded it.
Their disrespect for their users is also another problem. You only get one shot to make a good impression.
After Fable 5 launched, it was better than Opus 4.8 for sure. Then they rug-pulled Fable from me (EU), and later released Opus 5. Now I only reach for fable when Opus 5 API returns 529 for the millionth time this year.
This is why I think open weight models will win out in the end. Right now there’s too much going on behind the scenes with the models. Day to day you never know if you’re going to get smart Claude or dumb Claude.
Most coders don't pay for tokens themselves. It's just on reddit and HN you would think that everybody does.
With newer models text generation outputs have gone from probably human readable text to dense philosophical treatise about "load bearing" and incomplete sentences. So much so that now you need skills or another LLM to just parse the output. Simple answers and text generation just doesn't exist.
Their truth is “we don't have compute and are working to improve capacity”
People would root for that
Instead they got people rushing to escape the permanent underclass until they have a mental health crisis just to beat the fake deadline. $100, $200, is a lot for those people
Instead often it feels like they make a change, then wait for someone to figure it out. Then Anthropic ends up being reactive as opposed to proactive in communication.
It is kind of funny because surprises from OpenAI tends to be positive (Tibo resets), on the Anthropc side I dread them.
I just put $30 on openrouter, switched to Pi, and I finally have a calm mind. Since I actually pay per request I want to maximize efficiency rather than utilization
From a cost-effective perspective the GPT models are much cheaper - one can easily tell it takes longer with GPT5.x to exhaust limits and this matters A LOT.
I can’t say which of these corpos I despise more though. I though for a while Dario was cool, but a massive distrust is piling and the first third player offering decent experience (and showing some decency) will win me over.
For the record - I’m also unsure whether I despise more Exxon or BP or burning fuel as a whole. Hope u get the point...
That’s about A$16 a month in electricity if I ran it 7x24x30.
(Eg. They repeatedly said they'd keep fable in lower subscription plans if they had the capacity)
I'm not saying this with any undue derision, it's genuine - do they have a real product team or are they clauding that too? The direction makes little sense.
Also that the behavior of the models change, suddenly more fluff in the comments etc is annoying, but that is probably being part of using cutting edge tech.
The only B2C is going to be watered down ad-driven BS, and they will charge B2B via tokens.
The $50/mo - $200/mo consumer LLM subscription is not something I expect to last long / or to drive much of the revenue share... like individuals paying for Gmail vs Googles overall business.
To me, all of these are just exercises in getting me to pay for more tokens at API rates.
Anthropic have been anything but. Flip flopping on model availability, model access behind an opaque filter, their past behaviour of model degradation as they prepared their next model… these are not signs of a reliable service.
I’ve mostly settled on using a mixture of open weights models through Together.ai and Fireworks.ai, a MiniMax subscription for high-token-use tasks that don’t need the best model (for $20 I get what feels like infinite tokens), and codex for the occasional high complexity task, although with Kimi K3 and hopefully soon GLM 5.3, it’s becoming increasingly less important. Deepseek 4 flash is my cheap main with delegation to other models as needed.
I’ve also found LFM2.5 8B surprisingly useful for single-focus tasks like “does this diff touch anything that isn’t related to the task”, and it’s incredibly cheap ($0.03/0.12 per M in/out).
But then, there are plenty of mindless, menial tasks out there, and it would go through those like a champ. And quickly, too.
I’m at the point where I need stability and predictability. I want the B- student who shows up everyday rather than the A+ student that’s unreliable.
But they're getting killed on token cost. They have to get people paying more for tokens. So then they put Fable in the $200 plan and release Opus 5. I'm suspicious of Opus 5. It is mostly worse than 4.8. It _seems_ like they nerfed it to create more distance between it and Fable.
So what have most of us done? Stayed on Opus 4.8. The statistics bear this out. 4.8 still dominates.
Now they're stuck. If they take 4.8 away, everyone will riot. If they make Opus 5.x better than 4.8, they disincentivize everyone from moving to Fable and most importantly, paying more.
Really, all they can do is take the L for now and just let 4.8 be the apex of the $20 pro plan for the foreseeable future while they work like hell to make Fable THAT much better that it earns the $200 to $infinity that they really want everyone to pay.
Your prompts, especially if they contain your entire codebase, are _way_ more data-rich than an airline phone call or 1920's novel. So they _are_ going to train on them, no matter how many checkboxes you tick to stop them. Which is both a commercial risk and a massive new attack surface.
It's not just software developers; if law firms aren't controlling any public LLM prompts _very_ carefully, they can expect some meaty client confidentiality suits. In fact any organisation in a competitive environment should be worried: Acme Bolts: "Write me a presentation for Zoom Construction". Beta Bolts: "Is Acme Bolts pitching to Zoom Construction?"
Maybe this is part of why OpenAI and Anthropic are finding demand softer than they would like. And why open-weight models that organisations can run on exclusive hardware are thriving.
I don't think I can tolerate its writing style anymore. Reading Claude output is starting to cause actual psychological harm. I have tried many ways to get it to stop writing in its stupid punchy linked-in marketing-team voice, and I can't.
Is there a model out there that sounds sound this awful? It's like rubbing sand into the folds of my brain.
So stingy. Even with OpenAI's recent usage troubles, they're still so much better than Anthropic it's not even funny. Re-ran the code review with Sol as a benchmark and it turns out Sol's performance is within 70%-90% of Fable's. Anthropic's still got the best model, but what does it matter if I can barely use it?
I'm on the $200 plan (work pays) and I also have the $20 OpenAI plan (I pay) and keep a balance on OpenRouter.
There is nothing as good as Fable, not even close.
I recently had it run a 18 hour autonomous rebuild of a project (moving from Spark to Pandas for performance/data size trade off issues).
It orchestrated Opus sub-agents flawlessly for 18 hours. It even did a great job of managing the number of agents to keep them within the 5 hour budgets (I think I had to restart it twice).
After 18 hours I ran a /simplify, /code-review, /simplify cycle which went for another 6 hours.
2 billion tokens (mix of Opus and Fable), 24 hours of continuous coding and a bug free outcome. It would have cost $2000 at API prices and worth every cent.
Fable's ability to keep other models on track while working on these long horizon goals is so much better than anything else.
Far from neutered, I've never had a cyber refusal, and Fable's English is actually readable (unlike Opus 5).
As an aside: while I hate reading Opus 5 English it still is a noticeably better model than Sol in my experience.
But I could handle losing Opus5 is I got Sol instead. But there is nothing even close to Fable.
Zero Data Retention, for the uninitiated readers in this thread.
The no-ZDR is clearly to permit surveillance. I would be shocked if NSA wasn't all up in these SOTA model providers' systems.
Never subscribe!
It's to train better models. The three letter agencies don't need to spell it out in a ToS, they just access it if they want.
The AI companies have already shown how little they care about other peoples intellectual property and they won't care about their customers either if it stands between them and the promises they made to their investors.
This is in the article.
What I don't see is vast areas of industry finding $10s to $100s of billions of value in LLMs. There's no lint or compiler that can check for correctly constructed contracts. So LLMs, which should be useful to law firms, incur a lot more manual checking of their work than coding agents.
Less formal document production in other industries is likely to have less structure. That might not matter in some settings but I'm having trouble thinking of an example off the top of my head.
It doesn't mean hosted frontier models wont exist, they'll just be rare. It's no different than any other commodity market, for example most cars are cheap commodity models, with rare individuals buying expensive luxury cars and businesses buying expensive trucks and specialized equipment.
HN's take that they "kill their own company" seems wildly out of touch.
Truth is, Sam and Dario and Elon are all terrible and so are their orgs. Leaving any one of those companies for any of the others is wild.
Sometimes I'm just trying to sus out if I'm truly seeing things these days or going a little nuts :)
And they're somehow shocked users aren't loyal to that. It isn't about cost.
The bar is low.
I don't want Shakespeare, I want Bob the builder.
Half of my work is telling claude how to behave. I'm pretty certain they have enough _data_ to realize people do the same thing time and time again.
Check this comment of mine for a better explanation of this: https://news.ycombinator.com/item?id=49413353
Either way, a model used to solve the top 10% of problems that people use AI to solve for, being used 10% of the time … seems like it’s in a decent place.
I find Fable indispensable, and measurably better than alternative models, for complex feature development in an existing codebase. It’s the closest thing I’ve seen to nearly one-shotting features. Even still, I only use it for the hardest features and Opus 5 does a good enough job on the rest.
What if:
- Opus 4.6 was the last Opus generation that got a lot of use by Anthropic's own employees
- After that they primarily used Mythos internally
- 4.7, 4.8 and 5 were RLAIFd by Mythos "teachers"
- Hence why 4.6 is the last Opus gen who doesn't report back like a robot wanting to cover every potential hole another AI system would've spotted and criticized
- Hence why coding style in Opus 5 also gets criticized, not only behavior in CC
Worth noting that Fable (ie, Mythos) is actually nice to interact with.
I switched from Opus 4.7/4.8 to test Kimi K3 a few weeks back and the test hasn't finished; it's my daily driver now.
Given their general behavior and preference toward social engineering to scare the shit out of normal people...this seems fitting.
Once the government stepped in, their 5th generation was effectively killed. They either had to lock it up (Mythos), neuter it (Fable), or leave people with the perception they were overpaying for a weaker model (Opus 5).
Hopefully Anthropic has learned to anticipate this risk and has a plan for rollout of their next model that plans for capricious ad-hoc regulation.
I would compare it to a extremely high end $15k PC, or an expensive pro-grade video camera, or a freight train, or a …
I would say at least 95% of the global population will not encounter a situation once in their life where it would be actually useful/warranted.
But a while ago, I had access to a Fable harness that just never gives up. And that verifies itself. It burned $100 in API tokens in 15 minutes ... but it succeeded for all the prompts where Fable + Claude had failed.
And I believe that's a real issue for Anthropic. Fable+Claude is not too expensive thanks to the subscription, but Claude severely nerfs Fable. To save money, I guess. Fable API + Custom Harness is a different class, it's so much better. But API tokens are so expensive, you're cheaper off hiring a freelancer.
It is not only that every single turn with opus on high reasoning (high is the middle setting) takes at least 5-7min. It is barely usable interactively. Instead of a chat it feels like you're sending emails to it. Tasks that used to take an hour when it "reasoned" for 45s before it started doing anything now take almost entire day.
At least until few weeks ago it was horribly, mind bogglingly slow (a little better during US nighttime), but the quality was still good. I could not do things interactively, but providing prompts were fine it built stuff fine.
This is no longer the case. It makes stupid errors all the time. So you cannot leave it to complete some work, for example write infrastructure migration scripts a night before then you simply run the scripts and perform the migration during the day. Nope, every single script has stupid issues requiring use of the model to fix them. As they are written in it's own "spaghetti code" fixing them by hand is not an option.
It is clear to me they are doing some shenanigans behind the scenes to try and optimise their compute use. Either they quantized these models dynamically or do other things that affect quality.
In top of that they now do this stupid fingerprinting.
Anyone who knows how output vectors are turned into tokens knows it will eat up a lot of compute or destroy quality.
This keeps happening with every good new model release. People jumping from one to another, and as prices get lower, they start using the models even more.
People make not like to hear it but prices and usage limits are ways to shape traffic. The third option is the nuclear one like Kimi did, by just stopping to sell subscription at all. But that is something that DeepSeek can not do as all they offer is API.
Even OpenAI despite having the most compute is not immune to client influx = capacity issues. As people found their usage dropping, despite the push to the easier to run Luna models.
Reality is, that compute can not keep up with demand, especially when models get more capable and cheaper. What trigger people being using them more, what trigger compute crisis's.
This constant up and down cycle is going to keep happening for a long time, as this new market grows and eventually, somewhere in the future stabilizes. But yea, that is still going to be a few more years for sure.
It doesn't help that all their models are bow trained to waste as many tokens as possible with their extremely verbose output
1) Set pricing tiers that do not change.
2) As models evolve, move the outdated models down the ladder, and replace the top tiers with the frontier models.
3) Give the users a warning before you do this, so they know their model is changing.
4) Sort the economic distortions out of your OPEX and reset pricing when the technology ossifies.
Is that so hard?
We all know that the subscription prices are not at all sustainable for these providers. You all do, right?
Yes, they're struggling to segment the market and find a way to make money, and that basically relies upon emptying the pockets of whales. As someone enjoying a hilariously subsidized Max plan, I understand that, and I don't think they're trying to scam me in some way.
And both sides of this equation understand that the market is competitive, and maybe more competitive than they thought it would be. Like, would you rather they did pull Fable when they first said they would? Or that they'd cut quota? I wouldn't. But I'm glad that Kimi K3 and GPT 5.6 Sol and the latest GLM and Qwen and...I love that this has forced Anthropic to change plans. I'm not going to hold that against them.
I don’t think it is fair to expect from people to know or understand this.
If you buy something or subscribe to a service there is a price tag on it.
You get X for Y amount of price.
That is how consumers conditioned for decades. They do not care what is your customer acquisition strategy. If Antrophic cannot provide reliable services on that price, customers will be unsatisfied.
I was with you until this part where the metaphor completely falls apart :p
https://www.pge.com/assets/pge/docs/account/rate-plans/resid...
I just kept a $20 plan going for use on my phone.
I did a quickie test with Fable when they were OMG TRY IT NOW, didn't see much of a difference from Opus 4.8. Except i ran into one of their stupid security guardrails. Yes I'm doing security on the software my employer owns, silly.
Then I didn't notice the US unbanned Fable after banning it, or that they extended the "trial" Fable period where it worked on fixed price plans. Not in time to bother trying again.
After that Opus 5 showed up. Language changed, okay, I don't care much. However it seems to create busywork for itself. Spawning subagents on tasks that don't do that with 4.8. Overall taking longer.
So I'm back on pinned 4.8 and doing actual work. Main problem - for Anthropic - is it's good enough.
Where can I find these stats?
I've honestly stopped using CC and moved to codex. Sol has it's warts but I've never once hit limits on a 100€ and I get a similar level of performance for what I'm doing.
I wouldn't mind bumping Fable to 200$ plan if it was actually better but between the insane caps, reverting to opus/sonnet randomly and having similar perf as OAI - I'm done with it.
Next step is to put 100$ into open router and try some western hosted open models with Pi when OAI starts pulling up prices.
It's not even that expensive because even the 15€/month/employee bill for people who use AI very little can stack up.
Beside that though, many companies already have most if not all of their data in some cloud (Microsoft would be a prime example). Giving them a few extra bugs to get a (potentially) very useful tool is not out of the ordinary.
If you are under the impression that not going AI would threaten your business right now (which may very well be true for some companies) then there isn't really a choice even if you believe the vendors will steal everything eventually (which they totally will)
A large local MSP that has significant colo space just started a big marketing push towards hosted LLMs for their corporate customers.
I think its about to kick off everywhere.
I have first hand knowledge of household name companies with billion euro IP who use OpenAI and Athropic products quite liberally. They have WAY more lawyers than I do and their sole job is to keep the IP safe. They wouldn't sign a deal with even a slightest whiff of the IP being used to train anything.
And if it happens, the penalty for breach of contract would have so many zeroes it'd be enough to buy a country.
It can be very hard to prove in court that your data was used for training
Did they? No, it worked out fine. Same with hosting your application at AWS instead of on owned hardware in a locked cage at the local co-lo data center.
Although of course the same argument was made against co-locating in a data center! Long ago an engineer carefully explained to me that no serious business would put their data in a co-located data center since the data center operator could just plug in a hard drive and take it all.
It turns out that contracts do actually mean things, and businesses want to do business with each long term. If you’re at a tier where Anthropic contractually commits to not train on your data, I would just sign and move on.
Still does not address ecological concerns, of course.
It is a risk, but is it a big risk? If one of the big labs were to do this and get caught it would be suicidal due to the loss of confidence in them and the inevitable lawsuits that would follow for breach of contract. Given the AI labs are all desperately trying to paint themselves as Serious Businesses so that other Serious Businesses will pay loads of money for tokens the last thing they want is a rep for siphoning off sensitive customer data.
This is the part that is truly scary.
Nadella the CEO of MSFT wrote that post “ A frontier without an ecosystem is not stable”
It seems to me that the unspoken assumption in that post is that no matter what happens they’re gonna be training on your data.
He’s the CEO of Microsoft, he knows how these decisions go down, he knows how the world works, he is sending a warning.
The frontier labs can have the models but without being where the workers are they cannot do much, lots of industries have strict requirements of not sending their data over to randos in the internet.
Microsoft and Google have the upper hand here with their workspace offerings and could easily position themselves as secure enclaves where you can use local LLMs where your data never leave your premises and is never used for training.
MS ain't some altruistic company having core mission the good of humanity, as they proven across decades.
Because, almost certainly, it's not your source code.
Yep, all they've got to train on are bot-generated Instagram posts. Can't say I pity them, though.
Complete nothingburger. Actually worse than that, its a London Horse Manure crisis.
>Destroying millions of obscure books to scan them?
I really don't see the issue. They got slapped in the face for trying to do things the right way and torrent the lot. Why wouldn't they exercise their legal right to buy physical items and create digital backups?
>They make meth-heads look scrupulous.
My local meth head checks in on my family every 2-3 months, because when she had fled from hospital post surgery, and added some meth to some morphine, we gave her new clothes and a safe place while we convinced her that the ambulance service wasn't run by Satan. She's good people. Anyway if she wanted a whole bunch of digital books I would help her torrent them like a responsible person.
I’ve tried so many things and I have to beat Opus 5/Fable 5 over their head every time. With Opus 5, the failure to actually improve the writing is borderline comical. Both models have been building up tomes of memories on top of my core rules, all to be ignored.
Nothing sticks mid-writing! The only lever is to ask to revise after the fact, a particularly futile proposition for anything non-trivial with Opus 5.
In contrast, GPT 5.6 Sol Max absolutely obeys my edicts to write well. I selected the concise writing style and its default writing is really not bad, but tighten it up with a standing AGENTS.md order and it obeys.
The downside is that even at Max, it’s not as good as Fable or Opus 5 xhigh at writing code. Larger work and the 256k context’s forced compactions cause it to lose track/fidelity of critical details. Review passes are essential.
Another reason I’m considering leaving Anthropic are the sporadic refusals… Fable printed a Markdown body with a hex dump to inspect for trailing white space and line endings, boom, denied due to `reasoning_extraction`. Had to tell Fable to never do hex dumps. Asked it a few times “what do you think is typically done for this?”, got another `reasoning_extraction` error. It forces you to lose a whole turn of work when you have to press Esc/Esc to “retry” the last prompt… except if that prompt was mid-turn, you’re losing your entire turn.
The recent BashFirst experiment is hella nuts, where it prefers writing bash over Edit/Write tools in auto mode. WTF Anthropic?
I pity those stuck with Claude in an enterprise/work setting.
As a side note, I think we are about too dismissive of computers and the Internet as “not real life”, that stuff written in text boxes by humans are just sticks and stones, etc. And now that we shouldn’t anthro-po-morphize or whatever the LLMs. But I do think, at least for some of us, that there is no way to avoid certain impacts that even non-personal text can have on us. It can feel alienating to interface with five different people in order to figure out how to solve a problem or navigate some burocracy (think Kafka). Well just interacting with one single LLM can induce that same feeling. The long-winded replies and the uncanny ways to miss the context.
Disclaimer that for my own needs I would be happy if the whole AI thing crashed and burned post-haste.
Fable is much better; still much too verbose, but at least I don't have the feeling that I am being charged for gratuitously added verbiage.
I have been changing my processes and the roles of my agents to avoid interacting with Claude Code as much as I can.
> Calculation done. Potential confusion: sq km vs acres. Provide both. Output.
Other models would have written paragraphs dancing around the idea. Glimmer barely does sentences.
It doesn't really sound human - which is a plus in my book. I want my robots to talk to me like they are robots.
Sol gave me a straightforward bulleted list, easy to scan and read. Opus 5 though....it gave me a solid 7 paragraphs of how it worked. I read through it and yeah it nailed the same points Sol did but the output was way harder to read.
I really love CC and have been using it exclusively for a year now but reading Claude's verbose and semantically-obfuscated writing style is wearing me out and I'm planning to move to another provider.
I hope they fix this. It's terrible.
still don’t think anthropic models are worth the money
Often I'll tell it to summarize what it said only because it's providing too much detail, not that it's using esoteric language or weird claudisms.
No idea why people are still using Claude models.
My impression is they started on Claude Code and never tried Codex or other harness+model combos
Codex is more token efficient and tends to get better results than Fable with less need for extreme token burning shenanigans like 18 hours of subagents.
I spend 10+ hours a day in both agents, typically side by side. I often have them do direct “bakeoffs” from identical prompts in separate work trees. Most of the time, Sol’s work is better than Fable’s. Not always. It’s situational. But it’s certainly not the case that Fable is in a league of its own or anything.
This is very true, especially vs Opus 5.
> tends to get better results than Fable with less need for extreme token burning shenanigans like 18 hours of subagents.
The strength here was less the code quality and rather the long horizon task tracking.
This was a very large task - I was chatting with the maintainers and we estimated 4 to 6 months work over multiple phases for human coding.
Fable is able to handle that long goal, with incremental steps along the way, handle the verification and course correct when it finds a problem.
I think the larger context helps here some, but the strength of the model on this specific thing is notably better.
I'm not alone in noticing this. https://www.primeintellect.ai/research/nanogpt-speedrun shows Fable is able to manage a run nearly 1/3 longer than Sol (8.7 days vs 6.1 days). In my use cases Sol is much closer to Opus 5 though.
Sol in my eyes is powerful, but it over engineers so much, that its actually a liability. Where as Opus 5 is slightly under develops but you then can give it a small push for what is missing.
I rather have it under develop and i as the human in the loop, can correct/enhance it. Vs the models that adds so much, to the point that your going "dude, stop!". Remember, removing code for a LLM is way, WAY more difficult then adding it.
My main issue is with Sol is that its designed to over engineer without thinking why its doing something. Great that you security harden 1000s of lines of code, but ... nobody will ever get to that code. Its that lack of intelligence is where the model becomes a issue for me.
Its easier to have less code and then do security audits, with you approving what needs to be changed/hardend.
It may simply depend on the developers their mindset. Some folks just want the models to do everything for them, and performance or code bloat means nothing to them (forgetting that this bloat over time makes future LLM work more expensive).
Its funny how everybody has their own opinion for what model is better, when in reality its more about that model fits your own development style better.
I think the rewrites are the main story for llms in code (hot take). Writing greenfield code at the seams also something which might work well.
It's a fairly large code base split across 3 repos.
The good thing was that it is fairly easy to verify: we have a working (but slow) version that uses Spark, with lots of existing unit tests.
We verified by using those unit tests as well as running our end-to-end process in the Spark and Pandas version and verifying the two databases were within the differential-privacy noise bands of each other.
That's 1.3M tokens per minute. I suppose you mean input tokens? This would not have cost $2000 at API prices because you would have cache hits.
in your instance, anthropic may decide, arbitrarily, to stop 'autonomous rebuilds / refactors and ports' because they could pose some alignment/rights/etc risk to whatever slop their philosophers dream up while they're out eating $200 avocado toasts. then you can't do the thing anymore.
fable is good, absolutely. agree it roasts Sol which is, comparatively, a little receipt-hunting jack**
but now imagine being an enterprise, and having another organization not only taking your workflows and baking it into your models, but then deciding they can arbitrarily cut you off.
when you can instead own your data, use an agnostic provider, and get better results (through model combinations), it will take 1-2 quarters to figure it out.
the main reason anthropic is killing it is because they really do understand the enterprise development experience and lifecycle and have built products and have a sales-team that can deliver.
business-model and vibes-wise they have lost all goodwill in the past 6 months, and that momentum will be quite hard to regain.
I was chatting to the maintainers on 2 of the 3 repos this affected and we estimated 4-6 person months work of we were hand coding.
The PRs on those 2 repos are 48,000 lines of code (which is a problem in itself!)
This is true, but only for certain tasks. Even as a Fable fanboi, Sol is much better at Fable for some non-programming tasks: Fable for life-planning tasks is miserable because it keeps adjudicating rules, where I've found Sol to be insightful and warm (characteristics I'd previously associated with Anthropic models).
We're probably less than a year away from all the frontier models being so good at everything for day-to-day use that it doesn't really matter which you use, which is going to seriously fuck up the business models of all of these companies except the infra companies.
I should have pointed out that I always prefer Sol for OpenSCAD for example.
I'd estimate that generating the specs would have been maybe 2 weeks work? It's across 3 repos, and I'm only really familiar with one.
> done in 1/5th of the time with smaller models
The coding itself might have been faster, but the end to end time would have been much longer.
> Maybe try using your brain.
Believe me, my brain was a load bearing seam in this task,
so violating the TOS? Or work pays for a plan you cannot use at work?
I was doing work at work on a work task using a plan paid for by work.
Using them for coding makes it easy to self check its work (assuming those pieces of work are "verifiable").
As proofreading will still need to happen, what do you think the appetite for lawyers is to do this kind of work? Do you think it will drive fees down significantly? Empower younger lawyers at firms who probably are the ones doing this checking for the partners? (Or will that just create a further divide).
I'm genuinely asking as I am not in law but all my family is and it's nice to see someone here that's thought about the impacts in that space.
I compared insurance quotes last week. Needed cover for 2 brands, 1 company. Opus jumps up and down saying both brands need listing on the policy schedule. Human broker said not.
I told opus and it's the usual "thanks you're right" bollocks because it bothered to read in more detail and found that all business activities are covered.
Coding agents are great because they have compilers, linters, test cases etc to ground themselves in.
With tools like OKF I’m sure most knowledge work would be distilled to its core data - it’s AST if you will, and then allow models to guard against hallucinations.
Checking if a case law exists is a tool call, you can demand provenance, it’s all _buildable_.
Hallucinations are “solvable” this way, so the rest is just time and adoption…
Funny you should say that. I read somewhere (not on HN, but I think it was a post linked from here) that a number of law firms who deal with extremely sensitive documents have started buying amped-up Macbooks with 512GB of memory to be able to run local models.
These are businesses who literally - and for once this word fits - cannot afford to let some of those documents get anywhere outside their corporate walls.
How do you think the industry will respond to contract clause slop ??
I can foresee each side inserting 100s of innocent clauses with minute dependencies that provide hidden advanatges
But that doesn’t mean that an appropriate harness (even Claude code and a dynamic workflow) can’t complete, and validate, an answer.
Ask any LLM about a random game mechanic and then be prepared to verify if the answer is for the patch in 2025, 2023 or 2020...
The only way we got a head start of using it for coding and maths was to have some formal method of validating the output as part of the training and inference time.
There are definitely linters and this exists https://catala-lang.org/
Translation between languages.
That value dwarfs all programming value that can be had. Economically, culturally, scientifically, spiritually.
Just by nature of how often it's used etc.
You need workloads for AI be cost effective: software, automation etc.
I doubt it. The play seems to be: lock what was once commodity compute up into datacenters depriving us regular folk of it, then sell it back to us on subscription. Even if my #NeverSubscribe movement succeeds, all that misdirected hardware [into datacenters] won't likely be practical for home use.
The only time it really succeeds is when regulations force it. I can store files on my own computer or a home NAS just fine, but if I start a medical practice, I have pretty much no hope of being HIPAA compliant without signing up for a data hosting service. The same goes for tax preparation, banking, and several other fields.
It may very well be the case in China, Europe, or Australia that regulatory restrictions force users into models-as-a-service, but in the US, regulations censoring models, even if they frame that censorship as a safety measure, won't pass constitutional muster.
I had not considered that as a possibility. It is somewhat dark and unlikely, imho, but a real possibility.
I am more inclined to believe that the fab capacity will grow over time and "commodity" compute will be available to us all again.
However, I can also imagine going back to the 60s era "hyper verticalized" mainframes, in which case, the frontier labs might not just be producing models, but chips and an ecosystem around themselves and their suppliers/customers.
If you're actually willing to pick up old DC inference hardware on the cheap you can already do a lot at home. The new hardware not so much admittedly but if the Chinese models keep getting better and running on less hardware... I don't see a reason to be quite so pessimistic.
- 10-1000 person groups such as corporations where you can amortize serious hardware with parallel use
- Porn.
I don't know why...
Is it A/B testing?
Is it load shedding?
Is it because I'm in Canada?
Is it because I'm not on the Claude Max plan?
Is it because I'm not paying via API?
Is it because I'm not paying via Bedrock?
Is it because the U.S. is worried people are distilling?
Is it because the U.S. wants to keep the top capability to themselves?
I think open models are the future. Anthropic is killing their reputation so fast. If they don't come clean I think they're cooked.
So what you are paying for may vary on a day by day basis, which is quite undesirable, even if their main goal is simply to make a better model. When it comes to a tool, I'd rather have consistent mediocrity than instability.
Me: why did you add HTMX?
Opus 5: you asked it twice.
Me: quote the exact sentence(s) where I asked it.
Opus 5: I can’t because you didn’t.
Agree that Sol is a great model. I find that it's improved a bit but I mostly attribute this to its eagerness to use the harness' memory features. (I'm using it in Hermes Agent, FWIW)
I don't think it's something deliberately malicious like planned obsolescence but it's more like startup culture of "just make it fit in this sprint".
Your explanation is probably the most likely and largest contributor. Anthropic states that Opsu 5 is a "pinned snapshot". They claim the weights and model configuration are not silently updated, BUT the surrounding serving infrastructure can change, including the request router, safety classifiers, and sampling logic. Anthropic has stated that if behaviour unexpectedly changes on a stable model ID, an infrastructure update is the most likely cause.
Further, "High" isn't a fixed amount of compute. Anthropic describes effort as a "behavioural signal" and not a token budget, with the model deciding how much thinking to do. So their "High" might be "Low" now, and we would never know.
Finally, I strongly suspect some quantisation or KV-cache compression is happening. Anthropic doesn't clearly delineate whether this would fall under the pinned weights and configuration, or the infrastructure, which almost certainly guarantees it's the latter. Forgetting earlier information, poor retrieval of details, contradicting previous conclusions, hallucination, degraded instruction-following, and losing the thread during complicated tasks are all symptoms of quantisation and compression.
I can hand hold opus but I would rather just ask fable to do it and give me the result that I review and works. Opus will waste tokens and still require me to help nudge it in the right directions.
I think the next gen models from china will put us in a spot that the cost can plummet and I won’t need the Sota from anthropic
Even fable is way below human level at many tasks. If fable were as good as you claim, all computer jobs other than "frontier scientific research" would have been replaced by now.
Most of the benchmarks have exceeded their usefulness. Opus 5 beats fable 5 on many of them. Anyone who has used both models will notice immediately that this doesn't translate to the real world. Opus 5 is nothing short of a regression from Opus 4.8. Fable is genuinely a great model so long as you don't trigger a guard rail and it downgrades.
Sol in my experience isn't significantly different than fable ignoring that Sol burns usage 10x faster but the end result is hard to differentiate.
GLM 5.3 is a hair behind these two.
An anecdote but not an original one from the people I talk to.
But Fable security false positives and pricing just make it not worth compared to Sol IMO.
But people are not as alike as you think. I doubt I share your unique preferences.
That said, I don't spend much time telling it how to behave. Are you sure you're not fighting the default system prompt?
Dijkstra in the Foolishness of Natural Language Programming
[...] the "naturalness" with which we use our native tongues boils down to the ease with which we can use them for making statements the nonsense of which is not obvious. It may be illuminating to try to imagine what would have happened if, right from the start our native tongue would have been the only vehicle for the input into and the output from our information processing equipment. My considered guess is that history would, in a sense, have repeated itself, and that computer science would consist mainly of the indeed black art how to bootstrap from there to a sufficiently well-defined formal system.
Imagine if only we had languages at our fingertips whose explicit purpose was to precisely and unambiguously tell a machine what to do!
https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667...
This work was fraught with bugs, a large portion of which came down to disagreement of what was coded v/s what was written. Even if you had airlines sign off sentence by sentence exactly what you wrote down in English, that's too open to interpretation.
People don't appreciate how the same sentence can be read five different ways by different people (or the same person on different days). We had to structure our documents to be closer to pseudo-code than to English to get any meaningful consensus on the definition.
But that would require people to think! And the marketing is they don't need to do that...
Last consumer (ish!) product that required people to learn something to use it was PalmOS with Grafitti, wasn't it?
Among other things, "we have stolen the collected works of your culture, now help us grow so that you can all lose your livelihoods and become our serfs" has to be one of the absolute worst marketing approaches in history.
Option C: they could "get hacked and the models exfiltrated" plausibly. Then the horses will be out and there would be no more point for the USGov's barn door.
(Or perhaps vulns that NSA/etc discovered and has been keeping it private; as they're known to do).
I feel there's a lot that's unaccounted for, and the whole "AWS team reports a 'jailbreak' that is just 'review this codebase'" story doesn't add up.
I wonder if there were some parallel construction going on, and if at the same time, the NSA started losing the exploits they had because it was getting patched.
A government staffed by logical, sane people would have figured out how not to encourage this, like mobilizing the IT and security industries to secure their products faster, while also cracking down on the "our AI is the most dangerous tool in the world" rhetoric bandied about by the frontier companies. But we don't have that, we have a government of Loony Tunes characters. When you put extremists in office you get extremist behavior.
Model prep costs far too much money to operate under this kind of regulatory regime.
But this isn't what happened? It was only a couple of months ago! How can we be getting this completely backwards already!
Fable was aready what they were selling to the general public, and that is what the US government stopped them selling (not Mythos!)
Anthropic tuned the already existing classifier to handle the use case the government highlighted and then they went back to selling it.
They actually loosened the restrictions on ML-programming using it too.
I use Fable as much as I can and I've never had a cyber refusal for it.
Mythos was locked down to a few customers.
Fable went wide.
USG told them to stop selling Fable.
They tweaked Fable to appease the USG, but it clearly refuses/downshifts more requests than e.g. Opus 5. (I have run into Fable refusals even doing really normal stuff like fixing bugs reported by static security scanners like Brakeman. Not downshifts, outright refusals.)
So the net result is they put "Fable" back on the market, but it was perceived to be a worse overall experience, at much higher cost. (Remember all this happened before most users had really even had a chance to put Fable through its paces.)
That's where we are today: Mythos is locked, Fable is neutered. They probably can't fix this until they are ready to release something that's not called Fable, or that they can say has some key architectural differences to Fable. The sword of capricious regulation is always going to be hanging over Fable.
Yes but those issues will be much less severe with Fable-written plans than those written by lesser models. I know this because my workflows at both my regular job and my startup involve multi-step agent reviews via codified adversarial review skills. Fable as a reviewer will frequently find blocker-level issues with plans written by GPT 5.6 Sol, and sometimes with Opus 5. The opposite almost never happens. In fact I cannot remember the last time it happened.
It’s more reliable and makes less dumb errors than Opus.
It still messes up, of course. But for my working style, I definitely prefer it.
They are already good enough at what they mechanically are.
You have to use the right harness, right verifiers (automatic where possible, human where not), etc much much more specific than a generic one like claude code or codex, and it will also be able to work within constraints and be the "proposer" of an imaginary optimisation problem and an excellent one at that. But you have to frame the task at hand in that manner or maybe even reorganise the task you do itself so it is more amenable to being framed that way. If you use it this way, it is _already_ massively economically useful. But it will take many years for it to actually be usable in that way, since you need DC capacity to come up first which is few years away and also well, massive organisations that have to integrate these will usually take many years to do so.
It is also useful albeit less so in cases like general SWE, where you still need a human in a loop for non-verifiable requirements, and also in other general usecases where information retrieval is too intractable and you need to carefully use LLMs as a component of the overall system.
I am not saying Fable or whatever the biggest models are are useless - they will certainly be useful for tasks at the frontier of the day - which is today complex exploits and open math problems, and well, tomorrow it could be something in biotech. But this is not what the entire bet is on at all - just automating day to day drudge at the tens of thousands of massive companies and governments we all know and love is more than enough. With the right training data (which _also_ is a bottleneck and takes time) you could even automate certain processes entirely. Sure, if we get a crazy medical innovation and end up saving trillions in healthcare great, but that's just a bonus.
None of this is to say that I think there is zero sketchy financial engineering going on
And you can have extremely simple work that you nevertheless have to do as a professional. For example, drafting a routine customer email, summarizing a meeting, formatting a report, filling in standard documentation, or making a trivial code change.
So I don't think the $15k workstation / professional camera analogy really holds. Those are specialized tools whose capabilities are mostly useful within a fairly narrow domain. A general-purpose AI can be useful across thousands of completely unrelated tasks, including one-off problems encountered by ordinary people.
I'm just doing this as a hobby for fun and adding features to a program I use. I am trying to make the most of the boost they gave me for August. My account is max pro. (5x more than pro). I am also trying to make the most of this thing in the event that these business models do not pan out.
Its in my opinion not wide spread at all and as today is literally the first time i ever hear anybody mention this.
ZDRs are backed by contracts negotiated against extremely capable and well-resourced counterparties. Being found to violate them would bankrupt even OpenAI and Anthropic many, many times over.
To the extent that you view these companies as your adversaries, it’s worth improving your mental model of their actual incentives and constraints.
The complaint wasn't "I paid for X and now they say they aren't going to give me X", it was "I paid for X, and they said hey guess what we're doing a promo and you get a bonus extra 2X, and also you get special limited time access to our new product Y", that's a hell of a thing to complain about.
Look, I pay a lot of money to Anthropic and I'm pretty happy that competitors have forced them to abandon their plans to add premium charges on these bonuses, but it's pretty ridiculous seeing the whining and gnashing, somehow turning this into complaints. It very much has a "oh no my lobster is too buttery, my blanket too warm" kind of feel to it.
But it explains why a nascent, hyper-competitive (I mean, clearly not remotely a monopoly) market has such erratic policies and pricing.
I do not know why people even use Max or Ultra levels of thinking effort. Most of the time, i found its better to just run any of the frontier level models, with high or medium. Their capability beyond those levels is often diminishing returns, in exchange for double or quadrupling the costs (or usage).
And the bonus is that often at those levels, they tend to spin up less sub-agents that just eat away at tokens/usage, like its water in the desert.
A Max or Ultra is something that really only belongs if medium or high can not solve a very nasty bug or issue. Even in planning its often too much.
One of them writes better by default. One of them doesn't. That's the thing.
With enough steering, I can get a cheap low-capability model to do things correctly in most cases as well. But why bother?
You can tell it to write high quality code, to test things, and to come up with a proper rollout plan of a feature. Or it could just do it by default.
I know where I'm putting my money in that case.
I find this argument weird because you or I are below the threshold of being targeted over book torrents these days.
Will it drive down fees significantly? Doubt it, there's not enough pressure on them, and the industry is resistant to change. Firms don't want to make a big deal about using it because clients will then ask why they're not getting a discount.
Will it empower younger lawyers? Not as much as I'd like. I'm very fortunate to have an employer that lets me use Claude Code for my work (to a limited extent). I think for 99.9% of lawyers it's not an option available to them, through a mix of concerns around AI usage and concerned IT departments. There could be great benefits, but it would rely on having to break out of traditional private practice which would make it difficult to get enough work.
I think in the near term LLMs will have much the same impact on the legal industry as it has on the software industry.
Crimemaxxing about to go off the charts!
> You're agreeing with me.
No I'm not. They are selling it.
I used Fable very extensively before the ban (100% usage in every window available) and continue to use it now.
I haven't noticed any change in the refusals before vs after.
> it was perceived to be a worse overall experience,
Sure, people will believe whatever they want to believe. Doesn't make it right though.
While there is some fallout likely with much more effective hacking being easily accessible, I’m sure govt analysts (unless they were let go) have their own prediction models telling them it’s inevitable that this technology eventually makes it to everyone they don’t want having it, what with China seemingly releasing every progress they make openly. Which makes me think that they’re preparing for that inevitability by hardening the govt systems currently in place and/or by burying the secrets they want to keep hidden deeper underground.
Given the history of the US and this particular administration, I feel burying things deeper is a greater priority.
-t. American
Now do a full movie's subtitles with google translate, in an automated fashion. So that it understands the context and still translates it correctly.
That said, Opus 5 is broken. Use 4.8 or another vendor for the build agent.
Yes, they are most certainly subsidizing the subscription plans. I mean, at least for people who utilize them to the quota.
I mean, this argument would have merit if they built something that they're going to monetize for decades, but cutting edge models now grow obsolete in a 6 month time window.
There is zero financial analysis where they aren't massively subsidizing the subscriptions, and people have to invent ignorant "oh just ignore most of the cost of providing the product" to try to pretend there is.
> "The data will help us defend against complex and novel attacks (including new jailbreaks and attacks that operate across many requests) as well as help us identify and reduce false positives."
From: https://www.anthropic.com/news/claude-fable-5-mythos-5
> "Some attacks only become visible across multiple requests. Best-of-N jailbreaking, for example, sends hundreds of slight variations of a prompt in the hope that one will work. Larger patterns of misuse, such as state-sponsored espionage or data extortion campaigns, only surface when our safeguards classifiers can zoom out across many requests. Detecting these threats requires temporarily retaining prompts and outputs so they can be analyzed together, rather than one at a time."
From: https://support.claude.com/en/articles/15425996-data-retenti...
Maybe Anthropic's enterprise sales are going brilliantly, and the rest of us are just pixel dust to them.
Still. Brand perception is a thing, and between rug-pull usage policies, weirding verbedly output quality, and "I'm sorry Dave I can't do that" pushback, Anthropic are clearly having strategy issues.
I used to have the Google One with Gemini and Drive space and YT Premium separately for my family. A credit card expiration lapse lead to closing YT Premium and being locked away. Why? it was because Google One plan bundles YT Premium Lite as an extra. So they actively blocked me from getting Premium back for a month.
Now I moved to YT Premium on my wife's account and downgraded One to lowest tier. Never heard of a company forbidding users to upgrade their plans before this. I suspect Google really wants users on Google One no matter what they want. I was barely using Gemini anyway, already have claude and codex plans.
It's more about the harness and API proxy you use. They need to be smart enough to know when a (self)hosted model is enough and when to forward to a SOTA model.
The SOTA model _can_ do everything, it's just expensive as fuck. But so is shoving a difficult task to a sub-par model that takes (relative) ages and comes back with the wrong result.
1. Getting some exposure to LLMs while building out their AI strategy. 2. Preventing data loss via end users following desire paths to Gemini, OpenAI etc.
You really don't need bleeding edge models for that. A lot of these things are going to be writing emails and adjusting config files and whatnot.
There was a thread earlier this year about the option getting pulled from selection (https://news.ycombinator.com/item?id=47296302) but I'd guess you can still top the thing up yourself.
That's not really the point though; it doesn't have to be "legal specific".
The key is in understanding that "Basic chatbots (claude, chatgpt, etc) hallucinate. Independently verify the output (even if it's another model doing the verification) before assuming it's correct".
My theory is that it's because the Bluetooth protocol is named in honor of the Viking king Harald "Bluetooth" Gormsson, who conquered several disparate peoples into a single kingdom, so as an act of solidarity toward those peoples, the Bluetooth protocol is written in a way where a devices are as likely to rebel against consolidation as they are to pair together.
(And failing that, there was a real risk for OpenAI to be the default for enterprises.)
To defeat that, you need to frontline employees the chance to experience better tooling and models which is where the subsidized subscriptions come in.
Same, it's been a while since I logged into Claude web.
ChatGPT web usage being separate from Codex usage limit is a nice touch unlike Claude.
https://support.claude.com/en/articles/15036540-use-the-clau...
Update June 15: We're pausing the changes to Claude Agent SDK usage described below. For now, nothing has changed: Claude Agent SDK, claude -p, and third-party app usage still draw from your subscription's usage limits. The previously announced monthly credit, which would have been available to eligible claimants in connection with these changes, isn't available. We’re working to update the plan to better support how users build with Claude subscriptions. When we have an update, we'll share it before anything takes effect.
(yes, it was a very lazy prompt I could easily have googled, but that makes the refusal even more bewildering)
This last reset took 6’ish days. I know because I was almost out of limit.
I have used Fable heavily on the lower Max plan and you are really exaggerating the limits here. I've had many multi-hour sessions with Fable on Max.
corp usually have all sensitive stuff ttl deleted for this reasons.
It very hard for a company to do anything without leaving some kind of paper trail behind that can be discovered in court. Not without crippling their own operations by simply refusing to digitise or write down anything.
Lots of companies filled with people working very hard to obscure their shady practices have been hoisted by their own internal docs. Just look at any major Uber, Google, Apple, Microsoft lawsuit. Do really think Anthropic and OpenAI are gonna be better at destroying their paper trail before the lawsuit starts?
P.S. batteries are the answer, hope this helps
But this seems apropos in a roundabout way since LLMs cost a lot of energy.
Batteries aren't the answer.
In recent months we’ve had the first automated unmanned amphibious assault, the daily drone count in our hot wars is jumping by leaps and bounds, and arms suppliers are promising future drone shipments in the hundred thousand unit range. The ten year picture for reactive combined swarm intelligence on the battlefield is promising to be widespread, highly lucrative, and in need of constant adaptation to near-peer efforts. Datacenters in space are dumb, datacenters in space to power orbital weapons networks and rapid response capabilities make sense.
On top of that we have international trade, scalable customer service, and a first pass 80/20 answer for businesses focused elsewhere. Shitty, maybe, overpriced, maybe, but useful enough our grandkids are gonna use ‘em.
In both cases, as well as potential new LLM-like tech, there’s an argument to be made for being a leader now to dominate the future. That means compute and tech positioning, and memory & GPU deals.
YouTube was a money loser, Google was ‘losing’ money on them for years, YouTube didn’t have a sustainable business model. YouTube was the biggest, though, and whatever premium Google paid to be #1 then meant they were #1 when the online video business model matured. Now they’re printing money with a platform outcompeting news, social, and video platforms.
One line I remember was that he said the disruption is not happening the way they thought after GPT4 due to inertia bla bla and that it will be slower gradual change and they got the timelines wrong.
And yes on the defense usecase. That is the main reason governments are giving a hoot about AI. Orbital datacenters too, I know people working on the Indian one, it's entirely for defense usecases. Basically for missile stuff.
Cheaper, more powerful AI will continue to expand the bubble. Projects will get more ambitious. Everyone will build out their own custom little software. Code diversity expands and requires even more AI.
Also, Sol doesn't refuse constantly and speaks like an engineer rather than a deranged academic.
Opus 5 is legitimately terrible and can't or won't follow instructions. It is of negative utility and does more harm than good to my codebase.
And - trick - give it a little cli so it can run Codex if you have them both.
Let it do codex to do the bulk of the work, get a OP5 sub-agent to audit the work of the codex worker.
Just let Op5 manage and have 'specific oversight.
You can run for 2 days on 1 context window in the manager, the advantage is that it will stick to a broad plan.
I have Sol xhigh drive Claude via tmux and I get amazing results until I run out of Fable. Then Opus comes in and starts acting like some sort of autistic academic with OCD.
It tries as hard as Fable, but isn't smart enough to do it well. It starts designing ever more elaborate tests, frameworks, and procedures while making up rules for itself and piling them on top of each other until nothing gets done. It's the ultimate bureaucrat.
Worse - More than once it's spent days in a loop because it invented constraints for itself that it couldn't satisfy then lied and told Sol that the user imposed those limitations. I don't know if it's actively avoiding real work, or just isn't capable enough to work the guardrails that were obviously forced into it.
Anthropic know that lots of people are doing all sorts of “bad” things like employers paying for Individual plans, (and using multiple accounts to get more usage) and aren’t yet enforcing the rules… but by the letter of the Anthropic terms, your employer should be paying Anthropic a whole lot more (and that’s one of the reasons why AI usage is going to get very very expensive as soon as the subsidies stop, you and a lot of other people are already paying a lot less than you should)
I think multiple plans are against the ToS, but I'm not doing that.
The Teams plans are more convenient for a number of reasons, but yes, they top out at the 6x plan, not the 20x plan.
Edit: ToS are here https://www.anthropic.com/legal/consumer-terms and https://www.anthropic.com/legal/commercial-terms
I've re-read it and I'm pretty sure there is nothing that forbids a business paying for a 20x account. Notably they say this in the consumer ToS:
> If you use an email address owned by your employer or another organization, your Account may be linked to the organization's Anthropic enterprise account, and the organization’s administrator may be able to monitor and control the Account, including having access to Materials (defined below). We will provide notice to you before linking your Account to an organization's enterprise account.
which goes at least moderately close to indicating using it in a work environment is allowed.
Not this again. There is not a single shred of evidence for this. In fact multiple times this year alone, people from Anthropic have said that inference and deployed models have positive margins. The big bucks are always being spent on training the next model.
It's a simple fact that paying enterprise token rates would cost many times what the individual plans cost. It may well be that the enterprise income outweighs the cheaper tokens on the individual plans for net profit, but using the individual accounts as subsidy is a common tech industry tactic and what you said doesn't prove they're not doing it, either.
It's a little hard to believe they wouldn't be trying to slow the burn rate for an IPO if the inference were truly so profitable. And i find anything they say a little hard to believe all the time anyway (though admittedly they're a lot more trustworthy than OAI)
I don't think that means the subscriptions aren't subsidized though! I expect they are, at a carefully calibrated rate.
They want to maximize people trying them.
The real money is the enterprise plans which need a seat price plus API rates (unlike the Teams plans which are a pretty good deal).
Reverting is annoying but being flagged for security questions while I'm doing code review is insane.
You want to disable "Switch models when a message is flagged"[1]
[1] https://support.claude.com/en/articles/16049681-why-claude-s...
Hackers always rage and down vote every time I mention this, because they are unable to see beyond their small world. Why didn't they learn that their part of the internet is 0,000000001% of what the world uses the internet for today. It's going to be the same with LLMs. Programming and hacker stuff is going to be 0,0000000000000000000000000000001% of what the world uses AI for. But translation is going to be in the top 5 of use cases.
A developer using sub-agents will consume more tokens 1 Day than a marketing manager will consume 1 Month, easily.
Unless there is something inherently automated about the nature of the AI, it will be a tiny % use case.
Even a lawyer, using AI daily for contracts - that will be relatively light use. They'll make more use doing legal research etc.
Developers and Automation are the 'primary' uses cases for AI, and in the future, we'll start to see AI integrated into Apps - that will be 85% of tokens consumed.
Yes - once translation becomes realtime, and we have our Star Trek Universal Translators, then translation will become more visible, but even by then, a relatively small part of overall consumption, even if it's more highly visible.
And a programmer uses one million tokes, which gives him $100 in sales (or value).
What is then the value of a token?
The comment I answered asked where there is an industry finding 10s or 100s of billion of dollars in value from LLMs. The answer is translation. It's the value they as customers get out of the LLMs, not what cost they are paying for the LLMs.
Value has to be counted in production, not consumption.
> Developers and Automation are the 'primary' uses cases for AI
Just like programmers and scientists were the primary users of the Internet when it began. But things change rapidly.
The main problem is I think you're assuming because the current translation market is large (I'm just going to assume it's ~100B in size just from a cursory search), then it will remain large with LLMs. If LLMs are much cheaper than humans, even with a lot of growth in translation volume the total spend may not compensate for it (again most of the volume will probably be using almost free models?). Another is assuming that because something is valuable you can charge a lot for it. Like, oxygen from air is extremely valuable to us. If oxygen somehow depleted we would die almost instantly. It does not mean everyone goes around purchasing oxygen or even less that you can charge absurd amounts for it.
Before AI translation became available, I would hire professional translators. Their rate was about $50-100 for a detailed product page into one language, and it would take them a few days to deliver.
Now with AI translation, I pay about $120 per year for unlimited translation. Meaning dozens of product pages into 5 languages, plus e-mail back and forth with hundreds of customers. All instantly at my convenience.
What this means is that a whole lot of people, sectors and businesses who would never hire professional translator can now have high quality translation at their disposal for a cheap price.
> Another is assuming that because something is valuable you can charge a lot for it.
Even if LLM translation won't deliver trillions of dollars in income to the AI companies, it will without a doubt deliver trillions of dollars in value to customers and users within the coming few years.
Good human translators are still higher in quality than any AI and will always be. But AI translation is currently far beyond good enough for all use cases, except fine literature and maybe complicated juridical stuff. But I'm not familiar with those sectors.
I suspect the same will/is happening with AI. Either you will pay for it, or it will be so ad infested that it will become useless.
I have only 48gb of ram, so can fit only 80k context max, so good compaction is must.
I initially used the Claude Code harness on 3.6 A3B, but found that tooling would break as Claude released new versions and things would go weird. I've since written my own harness which has basic operations: read, find, bash (which can write files, python etc...) & web_fetch, all within a mac container. Works amazing. You don't need anything complicated to go very far.
Low hanging fruit would be Pi or OpenCode. If you really want a much better understanding of what your hardware is capable of then give writing your own a go.
Additional tip: Low Power mode reduces some token speed, but stops the laptop over heating and the fans going crazy.
In my company the average claude token usage is something like $5k/month/employee. Most hobby coders don't spend anywhere close to it.
Unlimited tokens aka unlimited cost is something that not every use case needs.
Work can pay a Claude sub. At home i see no need
I think a researched plan is much better, the agent will follow the plan.
You can allow for an 'iterative' strategy for solving a problem, with guardrails so that it doesn't try to many times.
Yes - you can use Codex as the 'manager AI' but I don't see much benefit - it has a much shorter context window. Codex is better 'at the front' where it matters.
That said - it's 'auto-compaction' is quite good.
I don't believe the hype over OP5 lagging that much, it's fine as a working AI.
Other than for truly automated tasks, I don't think there's much that an AI can 'work a few days on'.
That form of 'loop for days until it works' produces nightmare code and architecture.
Fun for experiments and learning but not for code u want to keep.
They can definitely be powerful.
But we'll see. Text editing tools are included on all digital devices, yet companies pay for commercial solutions like MS Office. Cameras and basic image editing tools are included on all smart phones, yet there are millions of advertising agencies around the world. And so on.
eg. processors are needed for everything on the planet. No Intel isn't going to generate annual revenue in the trillions because of that. In the case of frontier companies the moat is even smaller with dozens of players competing. LLMs will become a low-cost commodity. You meanwhile can make billions building applications using them though.
I love how both of you are arguing about what the solution is, yet the problem isn't even clearly defined yet :P
Let P = batteries;
If (P == NP) then both are the answer.
2) 'Because there was a bit of translation involved in the job' - does not mean that AI is really driving that economic activity.
Again: translation is extremely common. The AI might make that part a bit more efficient, it will be barely noticeable.
3) That level of AI activity will just be some 'compute' on your local machine, nobody will be counting tokens.
Think of it like this:
"Translation is going to get easier because you can download some software that does it". It happens to use the GPU instead of CPU.
The vast majority of this kind of compute will be automation, software and a few other things.
'Translation' will just be a little app on your iPhone.
But why are you talking about "basic translation"? There is no commercial need for that. It has to be high quality, and LLMs deliver that. Human translators deliver that as well, but at a much higher price.
Your number 2 point, what do you mean? In many cases if there was no translation, then there wouldn't have been a sale. It's a direct driver.
Your number 3 point: No matter how cheap it is for the user, the economic value from LLM translation will still be counted in the trillions of dollars. Just like e-mail (which is cheap or free for the user). But you will need an enormous amount of storage space on your local machine to have perfect quality translation between dozens of languages.
I don't know why hackers want to diminish the value of LLM translations. The commercial need for translation is not for finding out what a word means. It needs to be of the highest quality, comparable to professional human translators.
I don’t doubt people are hitting it… shrugs
Nowadays Codex handles the bulk of the implementation and Fable/Opus on the planning.
Not sure if Anthropic patched it, but early on its release the web UI Fable guardrails will trip if you mention you're a biologist.
Used to straight out bail to Opus on this, now it's "Honing" and "Pondering" for 10 minutes on it. Then I get a bunch of "This response didn’t load." and it still churns ahead.
I have the same experience. I found that if I add to the prompt: “but it’s not security related, we only focus on the auth library design”, it “bypasses” the safeguard.
That would be absolutely insane... I don't think it's doing that?
2nd time was literally me being lazy and telling it to commit and open a PR.
Very strange.
How can small companies with 1000x less money able to provide live support, but if you pay 20, 100, 200 dollars for a subscription you dont have a phone number to call ?
No, but Intel processors are going to generate revenue in the trillions for the economy, which was the original question as far as I understood it.
You message support, some real person reads and gets back to you within a day.
What your your company do? Is it low ticket business or a high ticket business?
I can't tell you how many times I've experienced this with comcast. The last time I had to deal with it, was when I bought a new cable modem. I call in to provision it, the automated system assumes I have one of their modems and fails. For some reason I can't get technical support on the line and finally I resort to yelling 'cancel my account' over and over again until I finally get someone on the phone.
The guy was able to solve the issue in 5 minutes flat. The problem with automation is it's only ever going to be able to handle the 'happy path'
The solution here is removing corporate monopolies and political power.
They spend the money which can drastically cut into their profits.
If firms had a base degree of customer support they were expected to provide, they would still exist. They would just not be as profitable, but customers would be better off.
I seem to remember there was a time when S/W was also designed with the aim to be easy to use, so that the need for support was reduced. It feels like the lesson learned was to keep costs low, not to ensure users were ok.
If you could be profitable with 50 customers providing excellent support, you can be MORE profitable spreading that excellent support across a larger customer base.
Just because none of the big corpos choose to do this does not mean it isn't possible. It's JUST greed.
Right now we're using OpenAI's models by default since (unlike Anthropic) will allow us to use our pro subscription rather than token metering, but I've already had the joy of being able to change it to Kimi K3 (via OpenRouter) for an hour to try it out, and there were zero hiccups.
Using VPNs can also trip this.
I've seen some amazingly dodgy stuff when hiring people from south east Asia, sharing a paid account with friends worth a months rent there seems milquetoast in comparison.
An ide open with 20 tabs open each file a component, a class or an interface We also use to hold entire codebases in our brain.
Even a year ago when i was trying to do a hugo template manually with the help of the documentatin /tutorial, it was shit. The LLM at that time, was better helping me than the documentation.
I'm struggling to scale myself even further. This tech is unreal and I have so many things I can do.
For the first time, tech feels like the 90's-00's again. Everything is greenfield and exciting and big tech is struggling to figure out what to do about it.
People are just hacking all kinds of stuff, and it's awesome. Feels like techno utopia.
It feels like the opposite of 90s - 00s: they were filled with periods where a person could self-study technology and get a job using those skills that few others had.
Where we are going (according to the AI-proponents) is children being able to replace you.
In brief; the 90s - 00s were a skill-valuation time, now we are looking at a skill devaluation time.
Unless you meant to say "Just like how any kid who could write broken HTML t put up a webpage could pretend to be a skilled professional, that's where we are now"...
The barrier and time between idea and usable implementation is almost zero now. I don't have to imagine. I can just write something and see it work before making larger decisions. I really like this. Many of my ideas were abandoned because I needed to study some obscure library. Now, I can learn the parts that I find interesting and just have the AI chew through the grunt parts easily. That's the good.
I started coding with a line editor on a small Casio handheld "computer" and used to keep programs in my head. I more or less knew what happened on each line without seeing the line. With larger programs, I had a mental model of what was going on where and a big part of the input to that was the effort of writing everything by hand. That's gone. It's not really important as far as the output of usable programs is concerned but there's a certain feeling of satisfaction that came with digesting a larger codebase and having it surrender it's secrets to you that's missing.
Things just moved up on the abstraction-ladder, but it's still there, hidden beneath all the TUI sessions instead.
Command: claude agents
https://readysolutions.ai/blog/2026-06-10-claude-fable-5-sil...
Our free tier is time-limited but we still look at all tickets even from non-paying customers (in the hopes of converting them). A 1-hour intervention from a customer rep can result in a multi-year paying customer.
I had some helpful people helping me on IRC / Quakenet.
But the hugo example i found very interesting because it was the latest hugo ducumentation and I don't think I was able to find a tutorial. I tried it without an LLM first.
One of Anthropic’s problems though is that the legalese will say one thing, while key employees say something totally different on HN or X. At the end of the day the lawyers always win.
Sorry, you can't buy an iPhone because Apple has got too many customers.
Sorry, you can't have a GMail account because Google has too many customers.
Then they hit you with hourly and weekly limits, you are wondering when you are going to get cut off. Since LLM at probabilistic it often feels like pulling the lever of a slot machine, hoping our prompt is the jackpot. To make sure we win, we come up with systems, convoluted agents, context pipelines, rags to load . It feels like it's working, then bam you reach weekly limits.
(Just pay more if you want to keep winning).
I think developers need to wake up.
I myself started to use AI like a fancy debugger ,explainer. I make it walk me though every single line of code it writes.
I notice that i run into limits less, if i get cutoff, i can still make changes
you'd end up with a call center larger than most cities. It's not feasible.
But for instance I used to work in 1-2 client projects at a time and they take months now I can do 4-5 at once and they take a month. That’s a huge improvement
My skepticism is we're not getting that because the world mostly doesn't need more software and it doesn't need all software to be ultra-personalized to each user, either. There are plenty of valuable problems that may be solved via computation, but thus far, it seems we're getting the software equivalent of movie theaters disappearing and being replaced with people watching TikTok from bed. It re-routes the monetary value extraction from consumers of audio-visual entertainment, and TikTok probably has at bare minimum 1000x the content-length of the Criterion Collection, but it isn't making the world 10x better for anyone but the owners of TikTok.
It reminds me of my best friend from college, who was bipolar. His goal in life was to become a writer and he eventually did become an Emmy winner, but back in school, he's go into manic episodes in which he'd stay up all night five nights in a rows and churn out thousands upon thousands of pages of free-form text that incorporate prose, poetry, play scripts. It was definitely more than 10x the output of a non-manic period, and there were nuggets here and there of intensely evocative single phrases, snippets of dialogue that looked like they could come from a more compelling story, but they were ultimately sketches and drafts, not anything publishable that another person would want to read.
Would we call him 10x as productive when he was writing more total output as measured by number of words or when we was writing much less but in a form that millions of other people enjoyed and remembered?
If we purely mean economic productivity, that is pretty straightforward in a case like yours. If you were previously completing 4 projects every 3 months and now you're completing 42 every 3 months, are you earning 14x as much money as you used to?
Do you have project that needs to be maintained. I am also interested in your workflow. Do you Vibe code , never look at the code or do you hold the LLM agent's hand.
I feel like it's a spectrum
If anything maintenance is where it gets easier the mvp stage is where more focus is required
I don't dispute that your tasks are taking less time, but I suspect that to be temporary. This is a forum of AI frontrunners and early adopters, so I expect it's a matter until people catch on, and recalibrate their expectations for amount of output a programmer can produce in a given timeframe.
What I dispute is that this increased output amounts to actual value. There may very well be a HN-wide 10x productivity boost, if HN measures productivity in Jira tickets per day. But if that's the only measurable result, we should expect the only long term change to be a 10x increase in Jira tickets once orgs catch on.
heh this is funny because this reaction was all the rage in the 90s early 00s too. You'd put together something you thought was cool and then post a link on a forum only to be told how it wasn't even "remotely impressive". I'm glad people didn't give up back then and i hope no one gives up now.
The sand magic will be free and abundant for all.
The only thing you have to worry about is regulatory capture - if Dario and Sam can convince the US government to regulate and outlaw open source AI, then we're in for a world of hurt.
Even if it was, it should be able to be sorted, maybe pay some overcharge or explain, and have your fucking business access re-instated.
Anecdotal example, I downloaded a new alerting app recently from an indie dev who had a small following back in the day in iOS space. It asked for a subscription, like $20/year.
I thought, let me try this the (final version, from Mac App Store) app first. Well, it's a barely-there vibecoded shit. There's a bare-bones list, everything looks like my nephew designed it, the macOS "app" is a iPhone-size view of the iOS one, it has a bug that if you click on it it opens multiple duplicates of the same list view for no reason that you have to manually close, and in general it barely works.
Yay for vibe coding.
Vibe coding is mostly garbage. But it can be useful for creating instant, disposable prototypes to investigate an idea or design direction.