They compare different reasoning levels.
https://artificialanalysis.ai/models/deepseek-v4-1-flash-101...
Scroll down to the cost per task pareto front chart which includes more difficult levels by default: https://artificialanalysis.ai/models/qwen4-max#intelligence-...
Qwen4-Max is not in a good spot on that chart. Astra xhigh and Fable 5.2 medium had a similar cost per task with higher intelligence ratings.
It's possible that they will run Qwen 4 Max at other reasoning levels in the future, or it's possible that they never will. Regardless, it's unlikely to be better at any reasoning level on a price/performance basis than the GPT 6.1 Sol/GPT 6 Astra pareto frontier. It's close enough to Claude Opus 5.5 line at lower reasoning levels that it could potentially cross that if it doesn't loose too much performance at lower reasoning levels. But it's far enough from the global pareto frontier that it's unlikely that it would cross that at any resoning level.
It does not (universally) hold for long running workloads -- which is what people are increasingly doing.
Leave a Sonnet on medium, it will complete the task, spend more time and possibly significantly more token money, vs. Opus 5.5 on xhigh.
I always use Max effort in both Opus and Fable. It's actually faster overall because tasks get done faster with fewer "corrections" from my end.
On GPT models, I'd sometimes dabble in xhigh/high but that's because OpenAI usage quota are too limited.
This is wrong.
To make a vacuum analogy, one brand's Low setting could be another's Medium setting.
To compare you either give them the same resources ( Watts or $ ) and see how much they can do, or you give them a task and then check how many resources they required to finish it.
The label on the knob is irrelevant.
It's very clear part of America's diesease is the idea that everything can equate to a dollar cost.
> If they are not testing Chinese models with less than max thinking, then we should only compare them to other models also at max thinking.
Another analogy for this are review sites overly focusing on reviewing all the SKUs of GPUs that cost $1k+ for an audience that, according to Steam's Hardware Survey, predominantly use midrange X060s-equivalent cards. Sure, it's fun to know whether or not you can enable PT on a 5090 but for most of us, covering 4070 TIs and below in-depth (meaning actually taking the card apart, seeing how good/bad the past jobs are, possibly buying multiples of the same SKUs to evaluate QC, etc) is much more useful because of the value prop.
Sure, that’s not in question. The discussion here is not about what to measure (we agree that benchmarking only max is suboptimal) but how to compare if you have a fixed set of measurements, outside of your control, which includes only “max” for model A but a wider range from model B.
The person I was responding to was claiming that the sensible thing to do, then, was to truncate the model B measurements to “only max” as well.
Nah, there are interesting conclusions to be had in comparing various model/effort level combinations, esp. when it comes to cost and time per task.
I have my own private benchmark suite; it involves algorithmic problems, the need for good visual understanding, and the ability to run long workloads unattended (as the agents need to run a few iterations, self-evaluate the results, and hand over the assignments when they feel they have succeeded). Chinese models do poorly on such problems, and running them at anything less than max efforts is pointless; they can't even hand over the assignments graded against a mechanical set of tests they themselves can execute. Hand that same task over to, say, Sonnet medium and you have a weird outcome: excellent result, but way more time and tokens burned than, say, Opus 5.5 or Sol 6.1 running at high. I do those benchmarks just to be able to spot such patterns and then better choose the models for my real work - but i wouldn't have caught it if I wasnt running various model/effort combos.
I'd love a comparison on for example monthly quota usage of all ~$20/month subscriptions.
It's like arguing windows 11 just needs more bloatware.
These models don't have to all in one every task in the world.
They have altered their benchmark multiple times in the last 4 weeks (v5 in https://artificialanalysis.ai/articles/artificial-analysis-i...) and the last 5.0.1 change https://artificialanalysis.ai/methodology/intelligence-bench... pushed up Anthropic's models a bit, you can see that v5.0.0 set on the first link Opus 5.5 has a lower score than in v5.0.1, and v5.0.1 was released just before Fable dropped.
Some part of me believes that the benchmark mix was put together to favor Anthropic's models.
But admitting their model is weaker is a no go. So conspiracy theory here we go
It's mainly just a composite of existing benchmarks. It would be very easy to benchmax if, say, AA shared with a preferred lab which benchmarks it plans to upweight/downweight in the next version. Just speaking hypothetically, of course.
Right now the big labs are buying benchmark tasks from "benchmark labs" (or whatever they call themselves) for absurd amounts of money, and making sure they perform well on them.
The incentives this generates are terrible.
The big AI labs will only buy benchmark tasks that the frontier models perform poorly on (because they have trouble finding any public ones). So the tasks are being mass-written now directly based on the models they're supposed to train. Obviously the math says you shouldn't do this. If there is a causal connection between the benchmark and the model, you lose the normal guarantee that improving benchmark scores will improve model.
In fact improving such benchmarks may make the model worse. To understand that let's just give a ridiculous example. I create tasks like "1 3 5 11 13 17, give next number". The next number is 301. These are the house numbers on some random street in Brussels, Belgium (I'm guessing the big jump is because they "merged" 2 streets, and just added 300 to the numbers of the second part).
If the model trains on this, performance will actually decrease because obviously this is not how you should reason.
BUT making tasks under this "impossible to reason towards" part is of course how you make tasks that might even be reasonable, and of course frontier models are going to suck on tasks like this. So they sell!
Rarely seems to make them look good whenever I drop in to the site. Far more expensive for little gain. Even their own effort levels look bad, xhigh is 50% more in cost than high for zero improvement across a range of tasks.
Also, a generic benchmark might not translate well into what everyone needs. I keep building my own test cases that test models for my own work specifics. Only that benchmark is useful to me. I don't need it to win a PhD math quiz.
Neither option is perfect.
A third option would be updating the benchmark every 6-12 months, but that has its challenges as well as you need to make sure that the benchmark covers things the model hasn't seen or is harder for the models to game.
Interestingly, Artificial Analysis does have a score that uses IRT, it's their Progress Index [2].
[1]: https://en.wikipedia.org/wiki/Item_response_theory
[2]: https://artificialanalysis.ai/methodology/progress-index
(not to say there is no value in "zero-trust" environments, but often pure-benchmaxxing is a deadend; some benchmaxxing is always involved and much less risky options than "train on the ZDR test env" usually exist)
You could run the benchmarks using downstream providers like AWS if they are available.
You could vary the parameters of the tests to both help with LLMs learning as is and to help obscure that this is a test. Especially if sending them through a thirdparty, so the providers don't get identifying markers like IP addresses.
But ultimately it's another arms race like with viruses and anti-virus software.
Otherwise you have to rely completely on trust that AA is really measuring intelligence.
This is a permanent problem in benchmarking really. Transparency is required for trust but invites over-tuning.
I have been using Terra a lot and I'd say that most of the models (even small ones) are are roughly equivalent when it comes to working in a consistent, established codebase.
But in practice, Terra is not very good at staying on task compared to say Sol 6.1 or Sonnet 5.5. Terra 6.1 will often just pause, thinking the objective has been met or the turn has ended.
I can't remember the last time openAI was at the top of the leaderboard?
Google had a brief moment at the top recently but other than that I feel like it's always Anthropic.
For most things I'd use Kimi is probably better so the ranking is wrong for me.
However it also seems like in those benchmarks where Kimi underperformed it did do quite badly, whereas when Astra underperformed it did so less badly. How you reflect that into a single number I'm not sure - it's certainly difficult to get right and it doesn't seem obviously biased to me.
Certain labs have been shown to and will continue to train on public evals, whether unintentional or on purpose, any respectable ranking across multiple evals also needs private data to counteract this.
Do you know a better site?
You can easily come up with a difficult private benchmark and test them yourself to validate the intelligence gap. It is real.
China lost. There is no catching up. Ever. That ship sailed.
What’s the evidence for this? Are there plausible alternate explanations?
[1]: https://artificialanalysis.ai/models/qwen4-max