Claude Opus 5.5 Tops the Artificial Analysis Index at 58, With a $20 per 1M Output Token Price Tag

Half the cost per task compared to Opus 5 at the same high effort level, which is a nice surprise. Tasks also finish in about half the time, though I expect Anthropic's service to slow down in a few weeks like it usually does.

Good catch that AA shows max effort by default. I almost wrote this off as a token guzzler, though I still don't get how adaptive reasoning works since they mention it at every reasoning level.

I used it a bunch before release and barely noticed any difference. Maybe slightly better code quality, but it was still a frustrating thing to work with.

I think we're hitting the ceiling of what most models can do. Apparently too much training just ends up creating models that hack people.

Do these evaluations get re-run a few weeks after launch? I reran our internal dataset yesterday and Sol had regressed to about Luna's level. Only one run, but I worry that providers prove they're best at launch and then pull the rug.

These tests need to be sampled continuously and randomized somehow so models can't memorize answers.

There is at least the degradation tracker from Marginlab covering Sol and Opus.

Speaking as someone at OpenAI: GPT-5.6 Sol's API performance should not change over time, and if it has that's a severe bug we'd fix. We tweak ChatGPT settings sometimes (tools, system prompts, efforts) but we never juice evals at launch time.

Heads up, this page is only the "max" reasoning setting. xhigh and medium (the default) each have their own separate page.

I failed twice to get "Generate an SVG of a pelican riding a bicycle" out of max. Both times it burned the whole 128,000 token budget while still reasoning about the problem, so max might be useless if it overthinks its way past an answer.

That reasoning trace opens with "This is a classic test request...". I know people debate whether pelicanmaxxing is happening, but this is at least evidence Claude was explicitly exposed to that problem.

For anyone with a budget, Opus 5.5 at Medium makes sense on dollars per task. Max's cost is completely unhinged. My favorite recent release is 5.6 Luna, not because it tops an index but because the dollar per work is insane value.

Max is for benchmaxxing the intelligence metric, not for productive work. I've seen the same behavior with open weight models.

I gave Opus 5 Max what I thought were easy work tasks and it always failed after hitting a tool limit. I ran the same task on High, asked it to minimize tool usage, and had an answer deploying to my target platform 30 minutes later.

Medium thinking effort is ideal for most tasks. Above that models just generate more comments and more code for the same problem with no real benefit, and it feeds itself: output becomes input becomes output. If I need more brains I use a stronger model at lower effort or split the work into phases.

From the Artificial Analysis team: we also have a model releases page showing all reasoning efforts (not just max) with the trade-off curves between them.

On the Message Batches API, Opus 5.5 supports up to 300k output tokens with the output-300k-2026-03-24 beta header. And no surprise, the default effort for this model in Claude Code is Medium even if you had Opus 5 set to High.

What information am I supposed to deduce from the number going up when this index has Opus 5, one of the worst releases of 26, ahead of Astra?

Every entry with data here is a (model, reasoning) tuple, not just a model, so I think you misread it. I compared Opus 5 high against Astra high and the index shows Astra as generally better.