Claude Opus
by Anthropic · Model · Language & code models
strong
Still a top-three model for agents and coding, now behind Fable 5.2.
MAC Bench finds most gains from eight agents came from parallel exploration; Opus 5.5 was the only tested model with extra gain from coordination
AccessAPIPrice$4 in / $20 out per M tokens
Why we say so
Arena ranks Opus 5.5 second on agents and web development, and Artificial Analysis ties Fable 5.2 with it on its coding agent index. Bespoke Labs found it the only tested model that gained from coordinating multiple agents. HN commenters report its safeguards block biology and DNA work about half the time, and Gemini Neon is said to match or slightly beat it on Vals.
Versions
| Version | Released | Rank |
|---|---|---|
| Claude Opus 5.5LATEST strong Second on Arena agents and web development, with strong coding scores | Sep 24 | #2 image understanding#2 web development#3 agents |
| Claude Opus 5 | Jul 25 | #2 documents |
| Claude Opus 4.8 mixed Usable for reverse engineering but needs babysitting on harder targets | ||
| Claude Opus 4.7 | ||
| Claude Opus 4.6 | #3 search | |
| Claude Opus 4 |
Evidence
- Nov 19Main result first obtained by an internal Claude research model; public Opus 5.5 later reproduced it independently, per the Hexagon entry
- Nov 18Predicts the next amino acid of unseen proteins: 34.9% real vs 7.5% shuffled, a +27.4 point gap, suggesting protein-sequence training
- Nov 18Trails ProGen3-3B (+32.5 gap) on next-residue prediction
- Nov 18Roughly matches ProtGPT3-10B (+27.7 gap) on next-residue prediction
- Nov 15#2 on Arena agent
Show all 46
- Nov 15#2 on Arena webdev (1818)
- Nov 12Opus 5.5 at xhigh outscores Sonnet 5.5 at max with 40% lower cost per task: a higher tier beats more effort
- Nov 11#2 on Arena webdev (1815)
- Nov 11#3 on Arena text (1543)
- Nov 9#2 on Arena document (1509)
- Nov 9#3 on Arena search (1252)
- Nov 9#2 on Arena vision (1324)
- Oct 19#3 on Arena search (1251)
- Oct 11Said to be even better than GPT-6 Astra at reverse engineering binaries; no figures given
- Oct 8Took #1 on Arena agent
- Sep 27Claude Opus 5.5 released; frontier-level at 3D, design, coding and UI navigation but pricier than GPT-6 Astra
- Sep 27Presenter preferred Opus on piano, math video, Blender and Unreal tests; Astra on the water-balloon sim
- Sep 24Self-reported results beat Fable 5.1 and GPT-6 Astra on agentic coding, knowledge work and computer use
- Sep 24Ranked first on Vals Index; GPT-6 Astra fourth
- Sep 24Ranked first on KernelBench, well ahead of GPT-6 Astra
- Sep 24Lowest hallucination rate among Claude models but higher than GPT-6, Grok and Muse; about $6 per task, above Astra
- Sep 24Got 1 of 6 brain-tumor types right, tied with Astra, GLM 5.3 and Chimera 3 as the best so far
- Sep 24Fable-style guardrails: risky cyber, biology or distillation requests fall back to a weaker model
- Sep 23Took #1 on Arena webdev
- Sep 13One commenter finds Opus 5 more verbose and harder to read than Opus 4.6 and 4.8, which they prefer
- Sep 4Designed a textbook 74-series and GAL circuit for 640x480 VGA output from EEPROM; one uncaught error, fixed with a blue wire, then worked
- Sep 4Tops the 1 Sept EEBench leaderboard at 61.6%, ahead of Grok 4.6 (57.1%) and Claude Fable 5.1 (56.4%)
- Aug 31Model reliably reaches for `python -c` decoders, making it exploitable via Python's cwd-first import path
- Aug 28Top of Terminal-Bench-Science at 30.0% resolution with Claude Code, ahead of GPT Sol 22.4% and Fable 21.4%
- Aug 22Opus 5 (max effort) ranked second at 2,920 (53.6% gap closed) in 2.9 agent-days, ahead of Kimi K3 and GPT Sol
- Aug 21Commenters report Opus 5 is slower (a multi-step skill took 2 hours versus 20 minutes on older models) and worse than 4.8
- Aug 20Users report jargon-heavy, over-compressed prose; 'load-bearing' appears hundreds of times in one codebase's output
- Jul 27Anthropic's Claude Opus 5 claimed stronger reasoning and planning than Fable 5 on agentic coding; on all paid plans and API
- Jul 27Researched Q4 reports, then built a one-minute narrated financial video with HyperFrames and Gemini TTS from one prompt, self-verifying; cleaner than Kimi K3
- Jul 27Kebab-skewer HTML prompt took about 10 minutes and roughly $6-7; output beat Fable 5 and Opus 4.8 on flames and atmosphere
- Jul 27On DeepSWE, GPT Sol still ranks first; Anthropic's chart de-emphasised it. Opus 5 beats Fable on agentic coding but trails on legal and health
- Jul 27Ranks 2nd on Simple Bench behind Fable, 3rd on LiveBench behind GPT Sol and Fable, and 2nd on Vals Index just ahead of Kimi K3
- Jul 27From one prompt in Claude Code it built a working browser Windows 11 clone with Word, Excel, Store and games; PowerPoint was weak
- Jul 27Got 0 of 6 tumor scans right, while Chimera 3 got one; unlike Fable 5 it did not refuse the request
- Jul 27Strong at front end and 3D design, but roughly a tie with GPT 5.6 and Kimi K3 elsewhere; about twice as slow, much pricier
- Jul 27Anthropic reports over 30% on ARC-AGI-3, far above the next best model, GPT-5.6, at under 10%
- Jul 27Ranked first on Artificial Analysis Intelligence Index by one point over Fable 5, with no confidence intervals shown
- Jul 25Same kebab prompt: Opus 5 slightly ahead of Kimi K3 on flames and photorealism, Kimi's meat texture arguably better; subjective
- Mar 31Took #1 on Arena vision
- Mar 3Took #1 on Arena document
- Feb 25Took #1 on Arena search