TheoPhysEval was built from problems working physicists had not published, precisely so no model could have seen them. The best agent missed one, at an average cost of about $2.33 a problem.
Model
Claude Opus 5.5
Claude Opus 5.5 is an artificial intelligence model developed by Anthropic. This page lists what is publicly known about the model and every benchmark result Aivio News has recorded for it, each with its own source and date.
- Developer
- Anthropic
- Model
- Claude 5
- Released
- September 22, 2026
- Modalities
- Model, Image, Code
- Status
- Active
- Source
- anthropic.com
Benchmarks
All stories| Benchmark | Score | Source | Evaluated |
|---|---|---|---|
| Vibe Code Bench single agent, medium effort | 91.5% | vals.ai Independent Verified | — Source date: October 9, 2026 |
| Vibe Code Bench single agent, maximum effort | 89.8% | vals.ai Independent Verified | — Source date: October 9, 2026 |
More from Claude Opus 5.5
A sample of 250,000 conversations found 0.51% about elections and 0.002% breaking Anthropic's election rules. Testers found Claude misses misuse spread across steps.
Anthropic says it cut what distillers extract from Claude by 90 to 99%, and names Zhipu and Moonshot
The company says reasoning tokens pulled out by distillers are now 10 to 100 times lower than before, and that it has turned away trillions of tokens of paying traffic.
Artificial Analysis scores Muse Wonder Full at 56, fourth of 689 models. Meta sells the weaker Muse Wonder through its API and keeps the stronger one inside Muse Code.
Agent teams cost up to 5.1 times as much as one agent, and beat it in only one of four comparisons
On Vals AI's Vibe Code Bench, a lead agent delegating to subagents beat a single agent once out of four — and simply raising the single agent's reasoning effort did more for less.
Arena published an index built from 72,509 real agent sessions on the same day it announced a $200m round. OpenAI holds the first five places.
The company says its generally available models block most cyber work, and that vetted defenders can now apply for one of three graded reductions in that blocking.
Mistral's new trillion-parameter model does the security work Claude and GPT-6 refuse
ML4 is a 1 trillion-parameter open-weight model trained on 3,800 Nvidia chips in Mistral's own European data centres. The number the company leads with is 82% on a vulnerability test where, it says, Claude Opus 5.5 and GPT-6 Astra score near zero because they decline the task.
The Pentagon has stopped using Anthropic's tools, seven months after it blacklisted the company
Claude was still being used last week in intelligence work and in operations against Iran, people familiar with the matter said. Anthropic declined to comment.