Price per 1M tokens is meaningless(janilowski.pl) |
Price per 1M tokens is meaningless(janilowski.pl) |
The uncertainty of how to use this vastly vastly outweighs the price in a data centre - so buckle up, buy enoughbGPUs to experiment at a known cost and one day you will find the approach that gives you 10x returns - at that point pay any price per token but not till then
People don't like to hear this but the open models just aren't good for end to end agentic workflows.
There are some very very good small open models that can excel in certain finite bounded tasks, but the foundational models are essential to building out agentic pipelines that actually work.
This is (apparently) the conceit of SteveYegge / GasTown - no model can cope unassisted so chunk it up, run it and if it falls over remember the exact place and restart, merging it all in
But that’s not my point.
I believe that software is a new form of literacy and just as all Companies and societies are literate now, in the future (tm) companies will run exclusively on software - AI developed software and those who go all out will have the sort of advantages the Catholic Church had over .. guilds?
Anyhow, that’s me being AI optimist. But writing the code is going to be a small part of that transition - almost everything to do with LLMs that is claimed amazing (Computer vision is something else) - almost everything people say we need an LLM is stuff you could have done three years ago but your internal politics just would not let you. Oh look we can see if our policies are being met (you could have written the policies in code and solved the whole problem)
Im struggling to get it out but - almost everything AI is proposed for is stuff a well run engineering firm coukd have taken on. A software literate firm could have done without AI is where firms are hoping AI will Get them
Imagine how far ahead real software literate firms will be - as long as they don’t burn their runway in tokens
Which is why, the right play imo is still buy in-house as much as possible, engineer around the problems and explore the phase space at marginal Cost.
Then and only then think frontier models.
I'm still using Claude at work (they're the only approved provider), but wow are the smaller models starting to SMOKE the big ones. At this point, all I'd consider paying out of my own pocket for is the lowest-limit Anthropic/GPT plan to get a big model as the Steward, but I wouldn't pay for ANY of the Anthropic models as the workers who do all the work. And as time passes, I don't know if I'd even do that; the open models are serving SO well.
There are probably 100 competing versions what this phrase might encapsulate. Could you elaborate more on which version you are using exactly?
My experience is that frontier models are only marginally better and not close to the cost/value of the open models which are anywhere from 10-100x cheaper. Perhaps I'm not doing "end to end agentic workflows?"
Stuff like the latest DeepSeek, Kimchi and GLM are used and loved by many people. It's not using an open model that is difficult: it's having the hardware allowing to do so. It's pricey and require technical skills.
That's why most people who are using (excellent btw) open-weight models are just renting compute online.
Also risking it all for some distilled models is a recipe for disaster.
Love to hear more about how you structure the orchestrator etc
As for the Orchestrator, it's pretty simple. In essence, it's like "Jira/Trello/Kanban on autopilot". Work items have states, a state machine defines how those work items transition between states, states are todo, in progress, retrying, reviewing, code reviewing, done. work items also have connections, allowing the LLMs to specify a dependency graph, and the dependency graph informs the dispatch order/parallelism, as well as when branches have to be merged. I talk to the steward, the steward has tool calls for interacting with all the data, and the orchestrator auto-dispatches all the work that comes in. I can generate work as fast as I can describe it to the steward, and that's usually the bottleneck.
So far I haven't had to deal with "how do you get the LLM to re-organize the work mid flight due to a worker finding something not accounted for by the planning", but I assume it'll come soon. The most complicated digraph I've tossed at it was 9 items and 4 layers deep. The kind of work I've given it hasn't been scoped large enough yet, so we'll see how it tackles that.
Anthropic models also shut down on a lot of security-related work, which is what I've been spending a lot of time on lately. I expected Fable to refuse this kind of task, but even Opus 4.8 refuses to build a verification harness for security bugs, as that involves exercising a discovered bug to prove it's been fixed in an automated red/green way, which looks like exploit creation to Opus' guardrails. So, I have to use other models for that work, now, though most of the original benchmarks I built were built with Claude.
The LLM itself produces one token. Some tool adds that token to the input and runs it again, flogging the horse. Downstream another tool, some kind of harness, tries to control this stream by injecting tokens into the context and then sending it to the inference tool, and then trying to pattern-match the output.
Finally, there you are on CodePorn.yata paying for an agent to generate code, paying for an agent to tell you what's wrong with it, and paying for an agent to make it differently bad, and hopefully move on to the next task.
If it still hasn't dawned on you that this isn't just a bubble, but a snake-oil-bubble-bath, just try to imagine the paradigm shift whereby you go on github.com, assign an issue to an agent, the agent fixes it by rewriting the application in Pascal but a reviewing agent catches that you wanted it to print a measurement in Pascals (pa), and you don't pay for the work or the review, you only pay for work that one or two reviewing agents determine is up to par.
Nobody is going to do that because as soon as they test it they're going to have to do some math that won't make sense without admitting/realizing it's not some near-sentient, AGI rating 0.9 intelligence, it's just a text prediction algorithm that can pull out entire sentences when you use it to infer output on topics it trained on.
Careful you don't "it's just a text predictor" yourself into unemployment .
This downplays the incredible things that can be done with it.
There's a lot of noise, yes. How long has the web existed? And yet we're still figuring out how to optimize (HTTP/3).
Disregard the signal at your own expense.
Anthropic's own release announcement mentioned that it's less cost competitive per task than Opus at higher thinking levels. It's significantly cheaper at lower levels though.
I'm wondering if this is going to be a universal pattern of smaller models: they're less smart, so to achieve the same benchmark results they have to think a lot more and hence become expensive.
Benchmarks force models to solve the problem entirely by themselves, requiring thinking. But if you pair them with a smart model (who thinks and solves beforehand) they won't need to solve the hard parts and can run on low/med. I suspect that was Anthropic's intention.
They're selling "intelligence", automation, etc but if the service doesn't work as expected the user has to pay for that.
Also, LLMs don't "make mistakes". They don't think or act. Every single thing they output is a hallucination. It just happens that the vast majority of things align with reality.
Changing the fuel type, efficiency of your vehicle, driving distance, or driving conditions will all change how much it will cost you.
Fuel cost per unit volume does not become meaningless just because you are neglecting all of the other factors involved. That would be throwing away the only data point you have been using.
This is just asking for someone to amalgamate all of the factors involved into one simple, easy to game, index.
I can go to 10 different gas stations and buy the same amount of energy from them. When I put it in my car I’m going to get the same result out. The differences are very small.
Tokenizers aren't standardized to anywhere near that level. A "token" from one isn't the same as a token from another.
I want a model that generates commit messages fast. Currently I have to wait up to a minute or two. That model doesn’t need to score very highly on SWE benchmarks, just highly enough that it can write out a good enough message in a few seconds. If you tested it on ${current top tier benchmark} you’d think it’s way too costly when in fact it’s the best tradeoff.
I’ve wanted a fast model to generate commit messages. No idea what that would be, but it doesn’t have to pass the SWE benchmarks very well. Just well enough that it understands the codebase.
If the benchmarks are non-predictive, well, you can't use them for much of anything, which is of course a recurring problem with every benchmark ever.
Some models I tried (Mistral I think) had better tok/s, and roughly same billion parameters / scores on various benchmark... But they were _so_ verbose, that they generated many more tokens compared to a Qwen model of same caliber to answer the same thing.
So even though it had better generated tok/s, because so many more were generated, the clock time was longer.
And this compounds over mutli-turns: more generated token means more context used in the next turn (until some compaction or something runs)
For example there's some benchmarks that show that Opus for any task that requires a higher than `high` level of effort, may have actually been cheaper to use Fable on low even though the cost per token is drastically higher
Similarly with GPT 5.5 vs Opus. They simply look at the dollar amounts the labs assign to each model and run with it.
But part of the issue compounds on the fact that there are many people who simply default to the smartest model/effort and don't actually vary their model per task. So in some sense I don't actually blame them very much.
Well that's the problem with these black boxes. You really have no idea beforehand how many tokens a given task is going to take. There's simply too many variables involved. It's therefore only natural for people to assume "the cheaper and older model is probably going to cost less overall to use than the newer, more expensive one."
EDIT: this is like saying hourly rate or salary is meaningless. Different people have different output. You have to evaluate performance.
EDIT2: just pray the LLM providers don’t start taking Patrick McKenzie’s advice and start charging based on “value delivered”
We've started trying to do some comparison videos to capture more of the UX vs speed vs cost stuff e.g. https://www.linkedin.com/feed/update/urn:li:activity:7479891... which one of my team did for my LinkedIn account (disclaimer: marketing)
(In this particular case Deepseek was way slower than GPT 5.5 but I think that's because it installed Libreoffice half-way through the task!)
But I have a sinking feeling that many AI developers think “tokens” got their name from the same idea as “virtual tokens in a casino” which is more related to product pricing and business.
If yours is the only request in the batch it will cost them one full pass through the model.
If yours is one of 1024 inputs in the batch the per token cost is 1024x less.
The point at which the metrics become meaningless is when others become aware of them, and begin to optimise for them. Lines per code is is not a bad insight for development activity, only when the developers are not aware of the metric. Price per 1M tokens became meaningless when LLM providers started to optimise for it. It seems to be that Sonnet 5 is optimised to score well on AA intelligence whilst seemingly having a lower price per 1M tokens.
I think generally we are in an AI bubble, and it will at some point pop. The numbers simply don't make sense. I would gamble heavily on local cost per task to survive the LLM winter. Given that hardware is pretty much a fixed overhead, you probably want to optimise for task per kW - that's where I'm betting.
Other aspects are caching, often at 0.1X cost, where providers really differ in how efficient they are (Anthropic really good, Google not so much) and how chatty a model is (costing output tokens).
Pricing per token is at least reasonably straight forward. If you aren't getting value, you don't use the service. One doesn't buy a Ferrari and then complain that in their town Ferrari doesn't help them pick up women and hence it should cost less.
i've always wanted cost per prompt, but even that has too much variation.
I'm sure there are degenerate cases, but I'd bet a relatively small model could do the job.
(see their follow-up reply: "The cheapest-per-benchmark-task model would be useless to me if it cannot do the task I need.")
In either case, you need the right benchmark for the right task
Like, can someone who hangs around a site like HackerNews really say something like this with a straight face? About the AI progress of the past few years?
Laughable. Has the 24hr news cycle completely rotted our brains?
There's empirically someone up your comment saying that, better to address it with arguments rather than calling them brainrotted.