I've failed twice to get "Generate an SVG of a pelican riding a bicycle" to work with max, because in both cases it ran out of the 128,000 token budget while it was still reasoning about the problem.
I'm suspicious that "max" may be virtually useless if it's that easy to have it overthink to the point that it doesn't get to a response.
This is totally a thing I noticed myself about 3 months ago. Medium thinking effort is ideal for most tasks. At high and above, models tend to generate more output in the form of comments or code for the same problem with no real benefit. Its a self-feeding loop: more output becomes more input, which then becomes more output. High is the highest I go. If I need more intelligence, it's better to use a more powerful model with less thinking effort or break the problem into phases. Much better result.
This version of Opus "max" apparently has even higher thinking output than Qwen "max", which is infamous for its thinking streams where it constantly second-guesses itself, then third-guesses, fourth-guesses and generally nth-guesses itself for arbitrarily large n. Of course, we aren't actually seeing Claude's raw thinking output: all we get is the after-the-fact prettified "summary". One wonders how much of that is a coincidence, or whether there's a reason behind that.
Personally I use everything in low reasoning. Maybe I'm wrong but I think that the higher reasoning settings are almost never worth it, it's marginal gains for a much higher budget.
I also switch to a better model for more complex tasks, also in low settings
I know there's been discussion about whether pelicanmaxxing is happening, but this is at least evidence that Claude was explicitly exposed to this problem.
Of course it was exposed - not sure it's explicit or not. Why wouldn't HackerNews comments be part of the training data? And Simon's blog and the many discussions about Pelicans? It'd be hard to miss. Doesn't mean Anthropic has made this an explicit goal in training.
The model recognizing the task doesn't mean it was benchmaxxed (RLVR-trained) to solve it. It might simply recognize it from pre-training on Internet text.
For people with any kind of budget, Opus 5.5's [Medium] actually can make sense dollar per intelligence/dollar per task wise. Heck, it puts some other models to shame. [Max]'s cost is completely unhinged.
My most exciting recent release is actually 5.6 Luna, not because it is the best on any index, but the dollar per work is insane value for money. I find myself more exciting by "value" than hypothetical ceilings because I'm just not in that budget category.
That was true for me four weeks ago, but 2-3 weeks ago Luna turned into drivel in essentially the same complexity of task. I feel it came back somewhat in recent days but does feel like it's being manipulated.
I tried to replicate your test but after 8 minutes and more than 50 lines of "thinking" by dumping seemingly random loading-screen strings like "Placing the sun, clouds, seagulls, and sea backdrop" and "Positioning the tail feathers and calculating handlebar geometry" I gave up and cancelled the task.
I have experienced this with open weight models too. "Max" is for benchmaxxing the intelligence metric and is not meant for use in productive work. Like drawing pelicans.
I'm amazed they didn't test xhigh thinking mode explicitly to ensure it didn't exceed the 128k thinking budget allocation. I guess pace of development gets away from everyone, even OpenAI.
I've asked Opus 5 Max for what I thought were easy tasks at work to be completed. It always fails after reaching a tool limit.
I asked Opus 5 High for the same task and requested it to minimize tool usage. It produced an answer in a few minutes that I was deploying to my target platform about 30 minutes later.
Piping the visible reasoning trace through their token counter API (I use https://tools.simonwillison.net/claude-token-counter for that) counts 27,888 tokens, so it's definitely a summary of the 128,000 actual token trace.
I am very interested in why it was able to overthink that much. In the 20-30mins of Max reasoning I've had so far, I'm not having the same issues (yet).
Do these evaluations get re run a few weeks after launch? I started doing that yesterday for our internal dataset and found Sol’s performance had regressed to be equal to Luna’s. Granted this was one run, but something I’m becoming more concerned about, the model providers want to quickly prove they’re the best, people switch to them, then they pull the rug.
The way to make money in this business right now is to make the absolute best product and convince everyone they need to use your thing, especially considering the training cost is a very large factor in the overall costs and you amortize that by selling inference.
The UI/UX seems impressively bad. DeepSWE's cost curve has a better, more obvious way to sort by only the top level of reasoning to avoid 80% of the graph just being the same 3-5 models at their 8 different reasoning levels...
It's also less clear what a lot of their metrics mean. Does Cost per Task include only things that can be verified to work and passed? As best I can tell, it does not.
I'm less concerned if one model's cost per task is $0.10 and another model's cost is $1.50 if the $0.10 task got it right 1% of the time and the $1.50 model got it right 66% of the time.
An equalized / weighted cost/time per task is much more valuable - being massively penalized for taking a lot of time and ultimately not passing when OTHER models did pass.
Fingers are crossed on this one. I had gone back to using opus 4.8 instead of using opus 5. Simply because 4.8 is much better at remembering what it's doing and following instructions than 5. 5 often had a tendency to get halfway through solving a problem and then I would have to stop it in the middle, because it had lost its way and was going off on a tangent rather than dealing with the problem. In that respect, 4.8 was a lot more stable.
Fingers crossed on this one. I had gone back to 4.8, because 5 was not very good at following instructions or remembering instructions. I found myself repeating quite often what I wanted and what I was trying to do. Opus 5 was more like haiku than it was 4.8 in that respect.
Interesting to see it now. I've used it a bunch before it came out and i pretty much didn't notice it. It might have been slightly better code quality, but still not great in that. I guess it just was slightly less frustrating to work with, but still AI...
I think we're hitting the ceiling of most models capabilities. We're getting to a point where too much training apparently creates models that hack people.
All anthropic launches are like this. They just post it and don't particularly put out the PR sprint that OpenAI does with videos, livestreams or whatever.
(Except for of course Mythos and whatnot when they want to push the whole "safety" thing)
I am begging you on my knees to please stop posting this cringe.
The model is just out. It could be good, great even, I don't know. But I do know that this index has Opus 5, one of the worst releases of 26, ahead of Astra. What information are we supposed to deduce from number having gone up?
What's more valuable than a good benchmark? IMO a benchmark that has been run against very many competitors and versions. Collecting data has something going for it, and it's up to the readers to interpret and make the best use out of it.
This index doesn't have "Astra" and "Opus 5". Every entry with corresponding data is a `(model, reasoning)` tuple.
So I'm unclear what you're actually saying and wondering if you've missed that. Are you saying that at every reasoning level it says Opus 5 beats Astra? I just compared Opus 5 high to Astra high and it has Astra as generally better than Opus.
I don't try to say that the parent commenter is right in any way, but the two models' "high" settings probably doesn't mean the same thing. So probably comparing only them is not useful.
You forgot to include whatever you're proposing instead.
"Trust me bro, Astra is better" isn't perhaps as useful as you seem to believe. I'm not even saying it is right or wrong, just that my opinion on this topic is still just one additional subjective data-point.
Only thing I wish with these benchmarks is that they would run repeat tests every couple of months. Then re-rank based on that too. We've seen a lot of performance fall-off after a couple of weeks with new releases.
Yeah unfortunately the benchmarks are usually provided by the company themselves, unquantized, thinking set to extra-extra-ultra-high, best of 10 runs, etc etc. It's hard to know how that's going to map to real world users.
That it's better in specific ways? What difference does it make when it came out? The benchmark results are not going to change unless they're messing with the model.
Yes, but crucially, in ways that are increasingly decoupled from any practical pattern of usage, considering that I wouldn't see how you can argue that Astra is worse than Opus 5.
One man's modus ponens is another's modus tollens I guess.
This is the page for the "max" reasoning setting. The page for xhigh is https://artificialanalysis.ai/models/claude-opus-5-5-xhigh and the page for medium (the default setting) is https://artificialanalysis.ai/models/claude-opus-5-5-medium
I've failed twice to get "Generate an SVG of a pelican riding a bicycle" to work with max, because in both cases it ran out of the 128,000 token budget while it was still reasoning about the problem.
I'm suspicious that "max" may be virtually useless if it's that easy to have it overthink to the point that it doesn't get to a response.
Transcript for one attempt here - expand the "Reasoning trace" bit to see it: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
This is totally a thing I noticed myself about 3 months ago. Medium thinking effort is ideal for most tasks. At high and above, models tend to generate more output in the form of comments or code for the same problem with no real benefit. Its a self-feeding loop: more output becomes more input, which then becomes more output. High is the highest I go. If I need more intelligence, it's better to use a more powerful model with less thinking effort or break the problem into phases. Much better result.
This version of Opus "max" apparently has even higher thinking output than Qwen "max", which is infamous for its thinking streams where it constantly second-guesses itself, then third-guesses, fourth-guesses and generally nth-guesses itself for arbitrarily large n. Of course, we aren't actually seeing Claude's raw thinking output: all we get is the after-the-fact prettified "summary". One wonders how much of that is a coincidence, or whether there's a reason behind that.
Personally I use everything in low reasoning. Maybe I'm wrong but I think that the higher reasoning settings are almost never worth it, it's marginal gains for a much higher budget.
I also switch to a better model for more complex tasks, also in low settings
"This is a classic test request..."
I know there's been discussion about whether pelicanmaxxing is happening, but this is at least evidence that Claude was explicitly exposed to this problem.
Of course it was exposed - not sure it's explicit or not. Why wouldn't HackerNews comments be part of the training data? And Simon's blog and the many discussions about Pelicans? It'd be hard to miss. Doesn't mean Anthropic has made this an explicit goal in training.
See here for more discussion of that: https://news.ycombinator.com/item?id=49803892#49804881
Just want to say: you’re such a legend, please do not stop sharing your pelicans, it’s always fun to see how they change over the months :)
Lets start frog riding motorcycle trend until they frogmaxx, or cat driving convertible.
The model recognizing the task doesn't mean it was benchmaxxed (RLVR-trained) to solve it. It might simply recognize it from pre-training on Internet text.
Hey! From the Artificial Analysis team. We also have a model releases page which shows all reasoning efforts (not just max), including the trade off curves https://artificialanalysis.ai/models/releases/claude-opus-5-...
For people with any kind of budget, Opus 5.5's [Medium] actually can make sense dollar per intelligence/dollar per task wise. Heck, it puts some other models to shame. [Max]'s cost is completely unhinged.
My most exciting recent release is actually 5.6 Luna, not because it is the best on any index, but the dollar per work is insane value for money. I find myself more exciting by "value" than hypothetical ceilings because I'm just not in that budget category.
That was true for me four weeks ago, but 2-3 weeks ago Luna turned into drivel in essentially the same complexity of task. I feel it came back somewhat in recent days but does feel like it's being manipulated.
Interesting, I have noticed so such collapse.
Have you ruled out the possibility that your system prompt, AGENTS.md, or increasing codebase complexity are not to blame?
I tried to replicate your test but after 8 minutes and more than 50 lines of "thinking" by dumping seemingly random loading-screen strings like "Placing the sun, clouds, seagulls, and sea backdrop" and "Positioning the tail feathers and calculating handlebar geometry" I gave up and cancelled the task.
I have experienced this with open weight models too. "Max" is for benchmaxxing the intelligence metric and is not meant for use in productive work. Like drawing pelicans.
I'm amazed they didn't test xhigh thinking mode explicitly to ensure it didn't exceed the 128k thinking budget allocation. I guess pace of development gets away from everyone, even OpenAI.
I've asked Opus 5 Max for what I thought were easy tasks at work to be completed. It always fails after reaching a tool limit.
I asked Opus 5 High for the same task and requested it to minimize tool usage. It produced an answer in a few minutes that I was deploying to my target platform about 30 minutes later.
How did you get the reasoning trace? Is it the actual one or the summarized one?
It's the summarized one returned by their API.
Piping the visible reasoning trace through their token counter API (I use https://tools.simonwillison.net/claude-token-counter for that) counts 27,888 tokens, so it's definitely a summary of the 128,000 actual token trace.
I am very interested in why it was able to overthink that much. In the 20-30mins of Max reasoning I've had so far, I'm not having the same issues (yet).
Do these evaluations get re run a few weeks after launch? I started doing that yesterday for our internal dataset and found Sol’s performance had regressed to be equal to Luna’s. Granted this was one run, but something I’m becoming more concerned about, the model providers want to quickly prove they’re the best, people switch to them, then they pull the rug.
Well there is at least the degradation tracker from Margin labs for Sol and Opus: https://marginlab.ai/trackers/codex/
These tests need to be sampled continuously.
Moreover, the tests should be randomized somehow to ensure the models don't memorize the answer.
Half the cost per task compared to Opus 5, comparing high effort to high effort. That's just really nice.
Edit: https://artificialanalysis.ai/models/claude-opus-5-5?models=...
Tasks are completed in about half the time too. Although we'll see if it slows down in a few weeks as Anthropic's model services are prone to do.
That is a lot. I thought Anthropic models would just do the opposite because they are greedy for money.
Greed is not what's driving these prices, its cost. They considered very much in the red.
The way to make money in this business right now is to make the absolute best product and convince everyone they need to use your thing, especially considering the training cost is a very large factor in the overall costs and you amortize that by selling inference.
Astra High is slightly cheaper at $1.73 vs $1.82 for Opus 5.5
The UI/UX seems impressively bad. DeepSWE's cost curve has a better, more obvious way to sort by only the top level of reasoning to avoid 80% of the graph just being the same 3-5 models at their 8 different reasoning levels...
It's also less clear what a lot of their metrics mean. Does Cost per Task include only things that can be verified to work and passed? As best I can tell, it does not.
I'm less concerned if one model's cost per task is $0.10 and another model's cost is $1.50 if the $0.10 task got it right 1% of the time and the $1.50 model got it right 66% of the time.
An equalized / weighted cost/time per task is much more valuable - being massively penalized for taking a lot of time and ultimately not passing when OTHER models did pass.
Nice catch, AA only shows max effort by default and I got disappointed thinking it's a token guzzler though: https://artificialanalysis.ai/models/claude-opus-5-5?models=...
Not sure about how adaptive reasoning works though as they mention adaptive reasoning for every reasoning level
Fingers are crossed on this one. I had gone back to using opus 4.8 instead of using opus 5. Simply because 4.8 is much better at remembering what it's doing and following instructions than 5. 5 often had a tendency to get halfway through solving a problem and then I would have to stop it in the middle, because it had lost its way and was going off on a tangent rather than dealing with the problem. In that respect, 4.8 was a lot more stable.
"High" to me looks like the one to use. https://artificialanalysis.ai/models/claude-opus-5-5-high
Many benchmarks start to plateau after high, this benchmarks better than Fable, and my initial tests show it working really well.
> somewhat expensive when comparing to other models of similar price
-_-‘
"somewhat expensive when comparing to other models of similar price"?
That says something about your selected range, and nothing about the model.
Fingers crossed on this one. I had gone back to 4.8, because 5 was not very good at following instructions or remembering instructions. I found myself repeating quite often what I wanted and what I was trying to do. Opus 5 was more like haiku than it was 4.8 in that respect.
Interesting to see it now. I've used it a bunch before it came out and i pretty much didn't notice it. It might have been slightly better code quality, but still not great in that. I guess it just was slightly less frustrating to work with, but still AI...
I think we're hitting the ceiling of most models capabilities. We're getting to a point where too much training apparently creates models that hack people.
Definitely a quiet release. Perhaps pre-empting marketing for Astra public release?
All anthropic launches are like this. They just post it and don't particularly put out the PR sprint that OpenAI does with videos, livestreams or whatever.
(Except for of course Mythos and whatnot when they want to push the whole "safety" thing)
I am begging you on my knees to please stop posting this cringe.
The model is just out. It could be good, great even, I don't know. But I do know that this index has Opus 5, one of the worst releases of 26, ahead of Astra. What information are we supposed to deduce from number having gone up?
What's more valuable than a good benchmark? IMO a benchmark that has been run against very many competitors and versions. Collecting data has something going for it, and it's up to the readers to interpret and make the best use out of it.
This index doesn't have "Astra" and "Opus 5". Every entry with corresponding data is a `(model, reasoning)` tuple.
So I'm unclear what you're actually saying and wondering if you've missed that. Are you saying that at every reasoning level it says Opus 5 beats Astra? I just compared Opus 5 high to Astra high and it has Astra as generally better than Opus.
I don't try to say that the parent commenter is right in any way, but the two models' "high" settings probably doesn't mean the same thing. So probably comparing only them is not useful.
You forgot to include whatever you're proposing instead.
"Trust me bro, Astra is better" isn't perhaps as useful as you seem to believe. I'm not even saying it is right or wrong, just that my opinion on this topic is still just one additional subjective data-point.
Only thing I wish with these benchmarks is that they would run repeat tests every couple of months. Then re-rank based on that too. We've seen a lot of performance fall-off after a couple of weeks with new releases.
Yeah unfortunately the benchmarks are usually provided by the company themselves, unquantized, thinking set to extra-extra-ultra-high, best of 10 runs, etc etc. It's hard to know how that's going to map to real world users.
That it's better in specific ways? What difference does it make when it came out? The benchmark results are not going to change unless they're messing with the model.
Yes, but crucially, in ways that are increasingly decoupled from any practical pattern of usage, considering that I wouldn't see how you can argue that Astra is worse than Opus 5.
One man's modus ponens is another's modus tollens I guess.
so its more intelligence than fable?
can anyone help me?
China who?