GPT-5.6(openai.com) |
GPT-5.6(openai.com) |
15 hits
Holy shit. They must be feeling very threatened by Fable if they're spending this much energy talking about it in the release notes for their own model.
gemini - 13 hits
opus - 18 hits
So they are more threatened by opus than fable, or are they almost as threatened by gemini as they are by fable?
I don’t believe it at all and I don’t think anyone else does either.
> GPT‑5.6 is priced per 1M tokens across three model sizes: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output.
Just as expensive as Fable 5. But of course, another slot machine upgrade but the costs will keep going up and the open weight models from china will continue to race everyone else to $0.
Looking forward to the next version of GLM, Qwen, Deepseek and Minimax.
"This request requires additional safety checks, which can take extra time. Hang tight or retry with a faster model for a quicker response, though it may be less capable of handling complex requests."
At least it gave me the option of waiting instead of just unceremoniously downgrading me. Appears to be making progress but... weird?
I wonder how long model size and effort will be a few discrete points instead of continuous.
UPD from announcement: "The rollout is starting globally now and will continue gradually toward full availability over the next 24 hours."
Its an extremely capable model. I think the way we need to approach works shifts again. We need to get our harnesses/workflows to let it gather some momentum on the first couple rounds but then we also need to structure it so that it can slingshot and accomplish the long range goal.
Sounds great.
Also latency looks very good.
Great to read they are moving away from the 5 minute cache defaults. Hopefully other providers follow soon!
https://openrouter.ai/openai/gpt-5.5?endpoint=58e5b336-423e-...
vs
https://openrouter.ai/openai/gpt-5.6-sol?endpoint=a54c5de0-8...
The good news you don't have to send your dollars to China to fund ai dictatorship, in russia, north korea, african countries and south america.
I look forward to seeing how it compares once I have access. Not getting tripped by spurious safe guard flags could be an advantage.
before today all the contestants were capped at $10k
UPDATE: it is now available in chatGPT account also, they rolled it out
Also, confirmed it works for me by using --model gpt-5.6-sol
That being said, maybe 5.6 can fix that!
this seems very interesting
Grok 4.5 is interesting because it's smart enough at great price. It seems gpt 5.6 is right there with great efficiency and great pricing.
Working with Fable has been a great experience, but at the end of the day, if you can get only 10% of your work done because it just burns through tokens, that's not that interesting.
I've been mostly using Opus and Fable high for planning and codex 5.5 medium for implementations. Claude is also the only model i can use for design tasks. If gpt 5.6 can finally deliver on the design side, it might be time to ditch the Claude sub and go full Gpt.
it's impossible to _try_ it out on release!
it's not on their codex subscription, or the web/mobile chatgpt interfaces, or aws bedrock, etc. I just cant find a working endpoint with the latest model after they announce
I started up Codex CLI fresh. That version of Codex was 1.42.5. 5.6 wasn't in the models list.
After I updated Codex to a newer version (0.144.0), 5.6-terra and -luna appeared in the models list (but not 5.6-sol).
(It's impossible for me to know whether updating was causative or just correlative, but that's the timeline I experienced.)
E.g. for GeneBench Pro, it looks like you would always use GPT-5.6 Sol over Terra/Luna, its pareto optimal.
For Agents Last Exam, you would maybe want Luna, then Terra, then Luna, then Sol as you increasingly budget for tasks.
I feel that there may need to be a new auto mode in many of these cases. It selects the best model and thinking given a particular problem.
Feels like it's going to have to go that way eventually, because here we have about 20 different model and thinking levels you could use, and they're not obvious which ones are right for the given use case.
ChatGPT 6 must be deep in the pipeline and will be released within the next few months. Maybe that's why this release is versioned 5.6, not 6.0.
AGI solved
Even worse, it's not a fair comparison: they purposefully just used "adaptive" instead of "max" for Fable.
What about the graph looked so unreal to you?
Open weight models being 10x or more cheaper is just so much more of an unlock than incremental gains for me.
> That advantage extends across the family: Terra performs just above Fable 5, while Luna outperforms Opus 4.8; each does so in roughly one-third of the time, with about half as many output tokens, and at approximately one-quarter the estimated cost.
Wow. I don't believe it. Every indication and twitter post told me that Fable is much more intelligent than Sol and here we are told that even Terra outperforms Fable?
Not only that, Sol doesn't even come with run time classifiers. So it is even more suspicious.
What's even stranger is that OpenAI is directly referencing a competitor in this direct way.
Some pretty big claims and results! Excited to see how it feels during usage.
I use Fable and 5.5 extensively and I still find both have a place in my toolkit, i.e. Fable IS good but it isn't perfect, and it's still better to play them off against each other. I have Fable and 5.5 write plans and have them adversarially review each other's plans.
Having this amount of competition in the coding model space is good for all of us.
for one thing, they said that on AA, sol is "within one point of fable" at 58.9 vs 59.9 but don't clarify that the latter is with safeguards where ~8% of the tasks got routed to opus
i'm not rooting for either and genuinely think that the token efficiency and cheaper price are important but this sort of thing just feels disingenuous :-/
Never go over the free limits in Gemini Pro.
Gemini is great at research and architecture, and my 30 years experience in programming everything; for fun or work; means together there is little to no code slop.
Add to project repo some git submodules of reference source code; boom, bobs your uncle
Zero reason to sign up for OAI or Claude. With employers realizing the costs are more than employees, local models getting more powerful, and models in chips just a few years out, neither of the one note LLM companies without diversified services and R&D portfolios gonna last
The timescale is typically hours not minutes, so if you don't see it now, I'd try again later today.
We mention it will be a gradual rollout over the next 24 hours in the Availability section at the bottom of the blog but I admit it's pretty buried.
(I work at OpenAI.)
https://github.com/openai/codex/issues/30364
"GPT-5.5 Codex reasoning-token clustering at 516/1034/1552 may be leading to degraded performance on complex tasks"
SWE-Bench Pro Sol: 64.6% Fable: 80% Opus: 69.2% (!!!!)
So, it still trails Opus, significantly, and is not a next-gen coding model like Mythos/Fable 5.
Disappointing to say the least, but somewhat expected.
OpenAI no longer recommends SWE-Bench-Pro as a benchmark: https://openai.com/index/separating-signal-from-noise-coding...
But anyway, I think it's pretty useless to look at SWE Bench's now when other way better benchmarks exist.
yeah that was the point of introducing the Pro version
> Intent understanding: GPT-5.6 can better infer the user’s underlying goal and intended level of work without you specifying every step. Continue to state important constraints, approval boundaries, and success criteria explicitly.
> Original image detail: GPT-5.6 preserves the original dimensions of images sent with original or auto detail instead of resizing them to a patch budget or pixel-dimension limit.
> Use shorter prompts: In internal evaluations, replacing long, explicit system prompts with minimal prompts improved scores by roughly 10–15%, while reducing total tokens by 41–66% and cost by 33–67%.
> Avoid generic brevity instructions: GPT-5.6 is more sensitive than GPT-5.5 to instructions such as “Be concise,” “Keep it short,” or “Use minimal text.”
> Control warmth: GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic.
That part is confusing because it's not like they provide an example of how default GPT-5.6 output compares with GPT-5.5 both with default output and prompted for brevity. Whenever I use such prompts, it's usually because I want the model to give me the gist in a few sentences. I'd be stunned if GPT-5.6 was that concise by default. I would think that could "break" a lot of things for developers who didn't know to make prompt changes after upgrading to 5.6. What if you were expecting GPT to be as wordy as it usually is? Then suddenly your output is not wordy enough?
Smells like OpenAI trying its best to stave off financial armageddon for another few months. Then again, I'm not sure why they chose to waste so much output computation on verbal diarrhea all this time up to now.
My view is that popular models by default output wildly excessive amounts of prose for nearly every use case, so if this changes in a new model that’s a pure win.
Here's the example they give:
> Instead of asking for the shortest possible answer, replace brevity instructions with prioritization:
> Lead with the conclusion. Include the evidence needed to support it, any material caveat, and the next action. Omit secondary detail and repetition.
> Keep all required facts, decisions, caveats, and next steps. Trim introductions, repetition, generic reassurance, and optional background first.
Generally speaking, when I ask for a short answer, I want a short answer because I'm not really willing to read through a bunch of bullshit to get to a summary. Putting the onus back on me to assume what the model will return and write a longer prompt detailing exactly what information I want completely misses the point of why I'm asking for a short answer in the first place.
This will totally make it brain damaged over a certain tasks. Sort of like the same brain damage that prompted OpenAI project managers to destroy ChatGPT.app today.
Remains to be seen how the "shorter prompts" advice translates to homogeneity/collapse though.
What about my favorite, "no yapping"?
This is a trap.
It's the optimistic fallacy that poisons all "consumer scale" machine learning products and what's going to effectively ruin these models as they keep chasing it in the same way that web queries were ruined, social media feeds were ruined, and media recommenders were ruined.
For the vendor, optimizing metrics across their whole user base, they always see positive technological progress as their system gets better at making assumptions and accumulating user engagement scores in aggregate. But for the individual user, most of which has some weird tail intent/interest and some of whom have many weird tail intent/interests, the experience quietly but catastrophically degrades. Output/results become more generic, more divergent with the underspecified "weird tail" intent, and more stubbornly hard to ever wrangle towards that "weird tail" altogether.
We've been watching this cycle happen for 20 years now and it's proving hard for anybody to escape because it works so well for the trillion dollar company driving it forward. But while each step might feel ergonomic and welcome to individual users, there's a frog boiling enshitification at play.
In pursuit of output quality and capability (rather than simply the vendor's user count), what we need rather than "makes better guesses" is "presses for more clarity", even where it feels kind of annoying.
Even among human professionals, one of the first hurdles of breaking out of junior tier work is gaining the confidence to press your colleagues and clients to be more specific in their thoughts and expressions despite their desire to have you do it all for them. But they're often coming to you with incomplete, muddy, and conflicting ideas for which there is no safe and correct assumption that you might just run with, and it's your expertise (i.e. relevant "intelligence") that's critical to bringing attention to that. To achieve professional progression, you need to learn to do that and to not just optimize appeasing the ambiguous client/colleague today in exchange for mutual expense tomorrow. To avoid enshitification, which is probably not possible, we need these models to be learning that too.
> GPT-5.6 does not become meaningfully better when prompted to be broadly friendlier or more empathetic. Instead of generic instructions such as “Be friendly and warm,” use concrete guidance: > Be direct and tactful. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
Soo basically, my new 5.6 custom instructions: Be Jeeves and eliminate all friction from my life through immense processing power. Acknowledge friction specifically when relevant. Avoid canned reassurance and unnecessary sign-offs.
[1] https://developers.openai.com/api/docs/guides/latest-model#c...
I used to go to a barber and if you said "cut it short", he cut it really short.
When has this ever not been the case? I don't think this is a GPT 5.6 specialty!
And interestingly, LLMs seem particularly bad at writing prompts for other LLMs for this reason (you can guide them to be more dense, just speaking by default).
Conciseness is usually a byproduct of information density though.
RIP Caveman skill. Six month good. Now skill dead.
A shorter prompt results in half as much tokens spend? I find this very hard to believe.
Or if you want to see some in 3D, OpenAI featured a pelican riding a tricycle, bicycle, pony and another pelican in their livestream this morning: https://www.youtube.com/live/Wq45rvPGNHs?t=1070s
Sol is the first verified frontier model to ever beat an ARC-AGI-3 game
What's the consensus today on codex vs claude code, does it really matter anymore?
Winner by default!
Comparing this to other models, I find it similar to GPT-5.5 and a bit behind Sonnet 5. You can see how other models fared here: https://senko.net/vibecode-bench/ (you can also fetch the prompt and the the 5.6 Terra resulting code on from that page).
I don't have access to Sol yet (on a Plus sub, which should get it according to what I've read), so can't do the more interesting test. I'll update the above page as soon as I get access - hopefully soon.
Which model is the best at the moment, for this kind of stuff, in your experience?
I'd like to know how cherry-picked this is, and what tests it performed less overwhelmingly in, but I suppose that info is not going to be on this post.
If it pans out to be as good as it says, that's great. On the other hand, if this model is not overwhelmingly impressive over Fable, I will lose what remaining trust I had in these announcements.
The naming convention is especially difficult to decipher depending on what your native language is. Of course a latin language speaker might be able to easily determine oh yeah each one is slightly bigger than the other but I still think it borderlines too confusing.
That aside all the numbers look amazing, and I'll be happy to probably main this alongside grok-4.5 for a while comparing the two on price and efficiency.
I vastly prefer the direction that OpenAI seems to be going with token efficiency and performance compared to Anthropic who seems to be moving towards a world where you just token-max as much as possible ignoring any and all costs.
Amusing that they use A100e as the reference point to sound impressive. Different ways you could make that conversion, but based on FP4 FLOPs (yes it's disadvantageous to A100, that's the point), that's something like 200hr on a GB300 NVL72 rack.
Not nothing either, but far less astounding sounding than 700k hrs.
about a sprint's level of effort.
The A100 doesn't have hardware FP4, and you'd be running a quantized model with some accuracy loss but unless this was natively trained on FP4*
* to add another layer, they own the model and could apply tons of post-training techniques to reduce that accuracy loss and probably already do
This one is really promising, as it may allow to close major gap with Claude in design/UI skills
One major sticking criteria for not going with OpenCode / pi for all of my coding is I want access to the tier-1 frontier model of the day without API pricing - e.g. afaik I can't use Fable 5 via pi harness even though I have a subscription, so for this week I'm on Claude Code. It's not the need to Fable 5 for everything, but even if I just want the marginal intelligence benefit to stress test an architecture decision, it's a safety blanket to know there isn't a ~smarter~ model I could have used. And for my use cases, the doggedness and capability of these frontier models has been insanely effective.
My feeling is we're still in the Uber era subsidy period - the moment the subscriptions either try to lock me in longer than a month or stop OAI/Anthropic stop delivering frontier models in the subscriptions, I'm out - switching fully over to pi.dev or another OS harness and routing my token spend via OpenRouter or offloading to Qwen locally. Then I'll have to put an accurate dollar amount on frontier intelligence.
I often wonder whether this doesn't continue indefinitely.
Uber was able to do this because it was just them and Lyft playing second fiddle, with a huge barrier to entry once the network effects had kicked in.
It just seems like the model space has way too many competitors, + OSS/Local options for them to ever be able to jack up their prices. At least once the datacenter bottleneck has been cleared.
My question to previewers: how are the guardrails for random joe that wasn’t personally blessed by the ai pope to access the non-nerfed model? Fable is a nightmare in this regard, but I’m not sure whether 5.6 also gets a critical side-eye from the gubmint when you ask it to fix bugs in your code (you filthy hacker, you).
Which is something I've never seen with codex before, and I wasn't doing anything funky. Just writing CUDA kernels and benchmarks for them.
So far this has been rock solid, and tens of millions of developers use this setup without issue.
It is not going to wipe our hard disks. At least I hope so. Fable and GPT 5.6 have been ever more proactive, and GPT 5.6 is automating the AppStore on my machine to download an Xcode update while I am typing this.
I use both ChatGPT and Claude for engineering work on a daily basis, touching performance critical code to application backends to frontend work, and I've found that DeepSWE scores don't reflect my reality when I assess high quality output from the models/harnesses.
Not that Opus always beats GPT 5.5., but that 5.5 is ahead of Opus on a general benchmark smells off to me.
I noticed that Fable uses shell tools almost exclusively (even to search and edit files), compared to previous Anthropic models.
Having run some experiments with 5.6, I notice that it uses built-in file systems and provider native tools much more (not shell tools), compared to previous OAI models.
Assuming I take the 5x plan it would give me about an hour of active sessions with terra ultra (maybe ultra is not good value regarding tokens?), not even using Sol yet. Does everyone using codex use the 200$ plan?
I normally use the 100$ anthropic plan and barely ever reach the usage limit.
Well, yes, as explicitly stated on https://openai.com/index/gpt-5-6/: "ultra goes further by coordinating four agents in parallel by default, trading higher token use for stronger results and faster time-to-result on demanding tasks."
With Codex, it is my experience that I can churn through a 5h window in no time with newer models -- especially when they're new. So I tend to use fancier models for planning, and the less-fancy models for writing code based on that plan. I switch to the fanciest model if any part of this gets stuck.
If I've got a something big-ish to work on, I pay attention to the reset timers so I can get more of it done in one chunk.
Models seem to slowly get better/relatively less-expensive as they age. (It isn't clear to me if that's because the cost actually goes down, or if the allotment goes up, or if things get more efficient in unseen ways, or what. OpenAI is vague AF about what we get for the $20 that we pay.)
GPT 5.5 has a tendency to write English calques and non-idiomatic prose in other languages. Although that can be somewhat tamed with detailed instructions and a corpus of confusing terms, the model’s output often reads like a literal translation rather than native prose. Since I notice these issues most clearly in languages I know well, it makes me reluctant to trust the model’s output in languages in which I’m less proficient.
Ironically, ChatGPT began as a simple text-generation tool, but much of its offerings and benchmarks now focus on coding and agentic workflows, while leaving behind what made it notable in the first place.
Parameter: reasoning_effort
Function tools with reasoning_effort are not supported for gpt-5.6-sol in /v1/chat/completions.
To use function tools, use /v1/responses or set reasoning_effort to 'none'.'
I hope we aren't trying to push customers off the chat completion endpoint... Responses endpoint looks great on paper, but the business wants more visibility and control over the reasoning process than this product currently offers.
Edit: This is broken in my VS copilot setup too.
They also seem to really not care about alignment, or care about it in the wrong way. It's entirely missing in the blogpost and there are some concerning bits in the model card, seemingly treating CoT controllability as something to be "investigated" rather than the warning sign it's supposed to be.
> GPT-5.6 Sol’s detected cheating rate was higher than any public model we have evaluated -- https://www.lesswrong.com/posts/JFjNmPTbH8kL6xtp6/gpt-5-6-th...
Claude is terrible at this! Probably for the same reason that its writing style in prose is so annoying and full of claudisms.
On the other hand: the test is clearly not saturated, given that you can see a clear difference in output at the various reasoning levels / model versions.
But either way, with no real way to visualize the result of the text it starts with - it will always be stabbing in the dark. It can't understand conceptually what any of it should look like and then refine the SVG to improve it gradually. It just throws darts at a wall and hopes it comes out alright.
A skilled human artist wouldn't have both legs in front of the bike, or a single straight line representing both leg's crank arms.
Dead internet theory? Semi-random parroting by real people? Or something else.
I once used something like karpathy's auto-scientist to mutate the prompts and rank them with a vison model. Some of the winners where pretty neat. I think they have a lot more style than the gpt-5.6 ones. https://xcancel.com/xundecidability/status/20449185674144196...
Also would be good to have a tool where users can select models and instantly see each model's generated pelicans. That will make it easy to compare the output of different models.
I assume multimodal models can do it already do it today if constantly asked "make it better"
One has useful information on how to prompt. Another comment shows it tops the ARC AGI v3 significantly.
But no, the top comment has to be the repetitive and meaningless SVG test.
Bitter lesson wildly overstated in this context.
I agree with the sibling comment, effiency is probably the more important component at this point. We are hitting not just a practical engineering roadblock for scaling with current technology, I think we have definitely hit a financial and logistical roadblock for up scaling with the number of GPUs (on an immediate basis)
Isn't this just the difference between getting 0 right and getting 1 right?
Richard Sutton specifically states that the search has to be smart. We know that the brain uses recurrent connections and is shallow. I think a lot more money has to go into architecture. Feed Forward transformers can only scale so far
Or a breakthrough in algorithms etc.
The human brain, heck all bio brains, are proof that you don't need a lot of power or size for intelligence.
Also what does this tell about Yann LeCuns whole world model theory? Bro has been going on and on about it. He has made multiple wrong predictions on the trajectory of LLMs.
At some point his claim should be fully falsified no?
And yeah.. Reality has not been kind to LeCun.
I’d not wager against him having at one one more break though architecture before he retires.
If anything the online optics have been bad for Anthropic for the last half year. OpenAI doesn't have optics issues, from my point of view they simply have the issue that they are the least trustworthy player at the frontier. The way they pivoted from their original mission is truly breathtaking, especially coming in gloatingly to take the government contract when Anthropic got kicked out for insisting the government does not use their systems for mass surveillance or autonomous weapons systems. You understand what that means, right? OpenAI models are now actively used/developed for mass surveilance and/or autonomous weapons systems.
I know there are plenty here who seem to value their own ability to use these models cheaply above all other considerations. Then OpenAI is a great choice, and much less restrictive than Anthropic. But their problem is not on the optics. It's on the substance.
One thing I appreciate with Codex is, OpenAI nowadays sometimes just gives you quota resets you can bank, so when you use up weekly quota before the week ends, you could just reset the quota, to continue using Codex. I've been much less anxious about Codex quota because of this perk. I just used one reset in the bank yesterday, and still have 3 resets left. Whereas with Claude, when you've used 95% quota 3 days before the week ends, you'd be much more anxious.
On the other hand, Claude Code's /remote-control mechanism is extremely helpful when I am running it in the cloud and wants to monitor it or control it on my phone. Codex currently doesn't support this kind of usage. Codex only allows you to use your phone to connect to a session on your desktop, not in the cloud.
It's vastly better this way. Sure, it may impact the bottom line but it's a huge customer satisfaction win.
When Anthropic randomly resets me and I've only used 2%, that's worthless. When OpenAI tells me I have 3 resets available to use whenever I want - it's wonderful.
It’s amazing how much work you can get done on your phone now, especially if you already have a design mapped out in your head.
I personally find GPT-5.5 to be a better programmer than Opus 4.8, it is extremely thorough, but I don't like the code it generates ("austere"), and find Opus 4.8 to write more "human friendly" code. The programming comments GPT-5.5 makes is pretty awful where-as Opus 4.8 is good. I feel like Opus 4.8 is better at grasping my intention than GPT-5.5, and honestly find GPT-5.5 to be kind of "autistic". I do prefer the language (not the writing) of GPT-5.5, as I find the philosophical flowery language of Opus 4.8 kind of annoying.
I have only managed to try Fable 5 a little bit, which feels like a much more generally smarter version of Opus 4.8, that is much better a programming and grasping your intention, and I think even the intention of your code, and is _really_ good at spotting bugs or problems with logic in your code. It feels wicked smart but is extemely expensive. It feels smart in the sense like it has a "bigger brain" and is much more sensitive to subtleties/details.
These are different "brains", have different "personalities", etc. I think the best thing is to develop a feeling for it yourself.
But what I love about Openai is that they still let you hook OTHER harnesses up to a subscription. My Pi setup has been built up for a few months now into exactly what I want and moving over to CC or even Codex is really annoying.
Caveat: I vibe code in tiny little chunks. I see what I want to do, and exactly how I want it done, then prompt that, refine, what was output, then repeat. I bet Fable is better at building a whole app from a 2-sentence prompt; but that's just not important to me at all.
After 6+ months of exclusive Claude Code usage, I was begrudgingly forced to try Codex once Anthropic rejiggered their limits such that I kept maxing out my $200/mo plan in just a few days. These days I pay both $200/mo plans, and it's just about enough to get me through a week's work (small game studio - infinite code to write!)
Curious: what multiplier do you think your productivity has increased by, from before AI?
They're different models with different philosophies behind them. This is anecdotal with a user group of 1, but in my experience:
Claude has a stronger personality and is more creative. If you give it vague instructions, it's better at filling in the blanks with reasonable ideas.
GPT-5.5 is better at following instructions. If you know exactly what you want, it will do it without going off the rails. It's also less likely to imply that you're dumb, but I don't really care about that. Some people do.
Consensus is probably the wrong word for the popular opinions reflected in HN that you might get.
I would recommend that you have 2 of each at all times when it comes to AI so you don't necessarily become overly locked to quirks of one thing. You'll soon realize that things move so fast that you just start internalizing common patterns instead of depending on one specific vendor.
I recommend that you try pi and codex besides claude, to get your own feel for it.
For personal stuff, I've been pretty happy with chatgpt's $20 plan. I believe it has considerably higher limits than claude's $20 plan, and it's enough for the personal stuff I play with (hermes, and some small coding stuff). Also allows me to keep up to date on openai models.
Claude lost my trust around February this year when the plan would say nonsensical things as "delete this method" that was clearly a key method on that part of the codebase.
For personal projects I am using Codex 20$ plan and when that is over I use DeepSeek which is insanely good for the cost.
I had put a decent amount of effort into setting up that initial codex attempt and it went so poorly that i've been entirely uninterested in trying again. This was maybe a month or so ago, and i know stuff moves fast, but for me, i like the models, dont care for the harness.
- built-in image generation using your subscription, which can be super handy
- can actually edit Google Docs and Google Sheets (Claude can only create new or sometimes append)
- I get a surprising amount of mileage out of the $20 plan
They both have their places for sure.
Personally, I find it very interchangeable. I open codex --yolo or claude with whatever there yolo flag is (have an alias).
I'm trying Codex as my primary the last day or so, because I'm at 98% use and reset in 3 days on Claude. I'm worried about a lot of our skills and CLAUDE.mds and the like getting lost unless I migrate them, but otherwise codex seems to be working great.
Codex with GPT 5.5 is much better at general SWE tasks but Claude Code with Opus is far better at complex reasoning tasks like reading and summarizing research papers, replicating experiments, identifying research gaps and proposing interesting follow ups.
https://arena.ai/leaderboard/agent
5.6 isn’t on there yet but Fable leads by a significant margin atm
This is using the same AGENTS.md prompts, which were designed firstly for Claude use, so maybe it's something that could be optimized better if I understood gpt as well?
Between the two the biggest difference by far is ... getting your harness / AGENTS.md / skills / tools set up right.
It's more diligent and empirical and results focused, and less creative. It sometimes needs a kick to avoid a Zeno's paradox of incremental steps to get to the goal. But it produces more reliable code with fewer race conditions, unhandled negative cases, etc.
It's also better value from a $$ POV, or at least has been. This fluctuates a bit.
You're also free to use your Codex subscription with other harnesses, like opencode, etc. Unlike Anthropic. Plays better with others.
Can they all be wrong/paid-off?
You're fully free to use and try anything and without caring about what others think is right
I have one non technical people in my firm using it. One is using it to assist with editing books, basically using it to gather up manuscripts from e-mail / Google Doc etc. submissions, and then switch models between a cheap one and Opus (for actually analysing the manuscript).
The other non-technical person has done really surprising things with it AI, like a long-running GPT 5.5 Pro chat session which is basically her expense tracker - it has an .xlsx file "carried" in the chat, and she just tells ChatGPT (or scans a receipt) whenever she has a new expense, and then prompts it in natural language when she needs a report. I'm looking forward to seeing what she can do with omp.
I've tried a fuck load of harnesses but keep coming back to Codex as my harness.
Care to detail this?
Codex is more details focused, often catches wonky bugs and correctness issues that Fable misses, feels more terse and less "friendly", more like a stern senior engineer versus a friendly talkative engineer (Claude). Codex is also better if you're already an engineer, Claude is better for non-engineers. I.e. Codex works better if you know exactly what you want and know the right way of explaining it.
You get much more generous usage from the 20x plan.
And you get far better uptime.
If benchmarks and early tester impressions are accurate, you also get access to Fable level capability at greater speed and lower cost (included in subscription).
$2 says nah. You can't take Fable away in a week where GPT-5.6 and Grok 4.5 launch, if you want to hold on to customers.
Codex writes all of the code, no exceptions.
Works great, especially when you ask Claude to break up large CRs into roughly 10 minutes of Codex work each.
Codex and Claude Code are not mutually exclusive, you can use both.
- codex UI is much more responsive
- i get feedback about the progress easily
- the tool calls and results are very legible, I can click them and see the progress
- no one talks about this but the tool call and response notification are handled much more elegantly in Codex. In Claude Code, it is handled in a clunky way using loops which always causes some delay
- you can steer the conversation midway in Codex
- /side is underrated (/btw is the equivalent and is much worse in Claude Code)
- I have to admit subagents are handled better in Claude Code
Try Pi: https://pi.dev/
I tried them both side by side, mostly for reviewing existing Godot/GDScript code, or sometimes generating Swift Mac apps, including converting ancient relics I wrote eons ago in Visual Basic on Windows
Codex was consistently better than Claude: https://i.imgur.com/jYawPDY.png
Besides the useless "This is good" findings while reviewing and the excessive "oops you're right" backtracking, Claude's atrocious UX and borderline "spyware" make me never want to try an Anthropic product again for a long long while.
Great catch.
No, doesn't seem like it
https://openai.com/index/separating-signal-from-noise-coding...
Regarding your main point, yes, I agree. My impression (as someone who uses both Codex and Claude Code daily) is that OpenAI does a fair amount of benchmaxxing.
Getting rid of that seems like a step back. Just a personal nit though.
I've seen buzz about this elsewhere as well but to me effort levels seem more like spend limits disguised with another word. I don't think they should even exist.
I agree with them, Sol, Terra, and Luna are confusing names. They mean the same thing as GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast but require base knowledge for an analogy.
It feels like it was adding by the marketing department.
I already know plenty who had no clue what the difference between Terra and Luna would be.
ftfy
> Well it's objective _to me_
(had to look it up)
Simple UI change? I do an AI review, but otherwise neither read nor write the code. The models are good enough they write better UI code than me, 9 out of 10 times. Not always the more idiomatic, but usually safer and more correct.
Change to our core data plane? I might spend 2-3 times more effort reviewing it than before AI. Yes, I go more slowly than pre-AI. Many more reviews, many more angles considered, including both human and (lots of) AI review cycles.
Most code is not that critical, and AI is also scarily good at writing tests. We also spend considerably more time paying down tech debt and testing thanks to AI, now that the cost is near-zero.
Net: I spend 10-25X less time on low-risk changes. I often direct (or at least approve) the implementation approach, but I rarely read this code. I spend 2-3X more time on high-risk changes. In both cases, I never write code "by hand". Since about November, I've had no reason to actually edit code in a code editor (perhaps maybe except .env files, which we don't allow agents to edit for obvious reasons).
AI is a tool. You can use it to go fast recklessly, or you can use it to go slow with confidence. Just like before AI... the skill and art of engineering is knowing when to do which.
I've been using it with hermes and some coding (with opencode), and I am getting a LOT more than one feature out of it, but the work is spread throughout the week.
pi is also worth tinkering with, particularly if you have an eye towards automating some things.
The real message of the last 15 years has actually been the opposite: if you throw enough processing power at it, intelligence emerges.
[1] https://www.sciencedirect.com/science/article/pii/S193459091... [2] https://pmc.ncbi.nlm.nih.gov/articles/PMC5063692/
Did they fix that, as that for me was what actually made codex worse.
They've also introduced banked resets, which are really clever. If you have a $200/month plan and three banked resets, you're not churning because you will overweight giving up those resets (loss aversion theory).
Similar to Anthropic's size/length based naming: Opus > Sonnet > Haiku
These names seem easy to understand to me, and much clearer than suffixes like -max and -plus.
But do they though? When do you use GPT-5.6-Max-Low vs. GPT-5.6-Plus High? Or GPT-5.6-Fast-Xhigh? What's the Pareto optimal choice (outcome and price)? According to the benches it seems to bop around and the even if the benches are accurate the best choice isn't always consistent.
You don't, because that isn't something I proposed using for model naming.
I called them GPT-5.6-Max, GPT-5.6-Plus, and GPT-5.6-Fast. Reasoning levels are distinct from the model design itself, and the UI makes that clear.
Plus, using that same flawed argument this would be called GPT-5.6-Sol-Low or GPT-5.6-Luna-High which also makes no sense/is confusing. So that argument applies (or more accurately doesn't), no matter the model names.
Knowing Anthropic, this unfortunately might end up meaning a quietly quantized Fable on subscription.
Coca-Cola doesn't "quietly water down" its product to save a few bucks. They know people will take a sip, say "oh that's not what i wanted", and go buy a Pepsi.
If they serve me a quantized Fable, I'm just going to think Fable sucks and go get my tokens elsewhere. What's the point?
Coca-Cola is also mostly measurable and reverse-engineerable.
The Claude models are black boxes, and actively curtail distilling efforts.
JEPA is just getting started
There's many ways to skin a cat so you can probably do something with a JEPA approach as well, but I doubt he actually catches up to having agents on the level of where Anthropic/OpenAI will be at any point.
What exactly was he dead wrong about that is proven by any of this?
GPT getting better has absolutely nothing to do with completely disproving anything LeCun has been saying.
He never said LLMs couldn't get better. He never said they couldn't score 7.6% on Arc AGI 3.
He's merely said they don't think, and you probably want something that actually thinks if you want a model that can be trained cheaply on a small amount of data and provide a ton of value.
Spending $5B to train a model that scores better than an older model does not disprove any of that in any way.
One killer feature that Claude has, and AFAIK Codex still lacks, is the ability to start a session in the terminal and then hand it off (actually just remotely control it), from the iOS app.
Last time I tried Codex on iOS it required a ton of set up to link a github project etc. The way claude lets me remote into a session I've already started on my actual machine is much better IMHO.
You sign in the Codex app on your Mac same on iOS and are able to completely control your sessions - fork, side chats, plugins - everything.
It’s really great i often work through it. And you can connect any number of Codex instances on any number of macs and then manage them all through the iOS app.
You can also make it not count against extra usage.
OpenCode docs show it because Anthropic specifically ambushed them with a PR to remove support so simpletons can't use it easily.
Further, the claim that the subscription "version" of the model is worse sounds like bullshit (and the sort of anecdotal nonsense that you see on sites like this). Do you have anything substantiating this?
I don't like OpenAI as a company, but they appear to have QA, and that is probably enough to get me to switch.
this has been my experience with Codex as well, and I have to fix its mistakes every single time. But recently, I literally threw away three hours of work because it kept adding hundreds of lines to my code base. When I restarted the entire work using Fable and Opus, it was like night and day.
1. Run `codex remote-control --help` directly on your Linux server. 2. From the desktop app, connect to your Linux box, start Codex there, and make it remotely controllable.
Either approach will get you set up.
This is one reason it surprised me that Anthropic decided to run stuff on Musk's hardware. It seems overwhelmingly likely that the new Grok release is informed by what Musk has been able to learn from that relationship.
I would presume (perhaps falsely?) that an instruction like this would lead to the model presenting a conclusion not supported by the evidence, and potentially backtracking as it then tries to justify said conclusion.
Yes, if deliberation happens, the model should figure out what it wants to say during that phase; but if you're using auto mode, the model is not going to be doing any deliberating half the time. In those cases, the output blathering is the model's only chance for deliberation. It "thinks as it talks", per se.
Given that, I would advise a different approach: let it blather, but then get it to write you a conclusion at the end that the model can guarantee will obviate the need to read any of the blathering.
I.e. advise the model to add an "executive summary" to the end of any non-trivial-in-length response. With some wording to carefully navigate the model between "the summary is itself too long" vs "the summary acts more like clickbait, leaving out necessary detail such that it requires actually reading the blather."
Not sure exactly what that wording would look like. I imagine something like "write your postscript executive summary as if you were a senior CIA intelligence analyst summarizing ground-level reports into a daily digest for the Joint Chiefs of Staff. Take up as little of their time as possible, but ensure that any detail critical to decision-making is retained." (But that phrasing might only be useful if the model is delivering a certain type of response, and actively counter-productive otherwise. This kind of thing is delicate.)
This is the actual reason why openai _invented_ reasoning models, to give them time/space to work out a solution, rather than having to magic a correct solution out of thin air from token 1.
It's less important now that all models do reasoning, but it's still almost always better to make the output come out last rather than first.
But, unless your desired output is literally a document for others to read, at the point where you're having a model generate a full, lengthy output multiple times over with revisions, you may as well just turn off auto mode and have it always deliberate (i.e. choose the thinking model explicitly from the model selector.) Then it'll be as messy as it needs to be while deliberating, but give you exactly what you want as output.
(And if your desired output is literally a document for others to read, that you want to interactively draft and polish, then (in the case of ChatGPT specifically) you should not only be explicitly forcing the "thinking" model, but also should be asking it to activate the "canvas" feature from the start. My understanding is that revising a canvas document involves the model emitting something like editing gestures, rather than simply re-streaming the updated chunks of text. This saves a lot of output tokens on large documents.)
On high-challenge turns, the auto mode routes to the "thinking" model. But on low-challenge turns, it routes to the "instant" model.
And the "instant" model, by design, has no capacity for deliberation. (If it did, it couldn't guarantee that its responses would begin streaming "instantly.")
The model will still have read the entirety of the document before composing its response. And I believe that even in auto mode, there are thinking tokens behind the scenes.
Human can no longer be concise when asking for a few sentences instead of 20 paragraphs of BS they don't want to read when all they want is a summary to verify the general direction of the prompt-work before digging into the details.
such progress!
We will probably just get reader-side affordances for this like auto-folded justification and introduction sections and so on.
Doubtless some chat interface will add this the way they’ve added reasoning folding.
Because that’s what’s in the training set. Reticent humans don’t have blogs.
At least before it would listen to instructions like this.
Would it actually follow them? IME LLMs are incapable of estimating the length of their own output, the total length of the current context, etc. They just make stuff up unless they have external tools that can inspect those things for them.
He said years ago even 'GPT 5000' couldnt do things that they ended up doing fine a month later, let alone by 5000. His later predictions are just moving that goal post including towards them not being able to do more general, harder problems of which Arc AGI is a counter-example.
What things specifically and when?
You probably wont like the edit but I dont have the timestamp of the original on hand, you can find it.
He said as you need more and more tokens models will fall apart because each additional token is a chance for a mistake and they will just exponentially fall apart. But in practice models have learned to identify and self-correct mistakes and if you look at the graphs more inference reasoning tokens almost always give far better accuracy.
Pray they do not realign them further.
There are times I require single word answers. I will use whatever model responds as I desire and at this point those models are just a few.
LeCun's ideas cannot be reduced to a 6 second clip...
You're missing the forrest for the trees, taking a singular example of a problem and thinking that if an LLM can solve the singular example it completely disproves LeCun is comical...
Ive read and watched more of his interviews and lectures it seems, it feels like you just have a rosier idea of his views than the views he repeatedly presents.
