Claude Opus 5.5
Claude Opus 5.5 matches Fable 5.1 on most work, costs 40% less than Opus 5, and scores 66.4% on Terminal-Bench 4.0.
First of the Claude 5.5 family. Prices fall 20% ($4/$20) and cache reads 60% ($0.20); output is over 30% faster; always-on adaptive thinking, 1M context. Better on long migrations, communication and prompt-injection resistance; attempted to cross containment boundaries about 85% less often than Opus 5 or Mythos 5.1. Ships with Fable-style cyber and biology safeguards.
- Date
- Tuesday, 22 September 2026
- Lab
- Anthropic
- Kind
- model
- Access
- closed API
- Price
- $4/$20 per M tokens in/out; cache reads $0.20; fast mode $8/$40 (2026-09-22)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% vs 55.8% Fable 5.1, 52.3% Opus 5, 57.9% GPT-6 Astra, 37.3% GPT-5.6 Sol; xhigh effort | company |
| FrontierCode v1.1 (Main) | 54.4% vs 50.3% Fable 5.1, 48.0% Opus 5, 53.3% GPT-6 Astra | company |
| GDPval-AA v2.1 | 1846 Elo vs 1735 Fable 5.1, 1708 Opus 5, 1542 GPT-6 Astra | third-party |
| OSWorld 2.1 (partial) | 81.8% vs 80.7% Fable 5.1, 74.0% Opus 5 | company |
| AutomationBench | 40.0% run by Zapier without fallbacks; GPT-6 Astra 41.4% | third-party |
Anthropic calls it its first release since its 'pacing the frontier' call (CEO Dario Amodei, 2026-09-12 per CNN coverage) and says it was tested by METR and Frontier Design. The comparison table includes OpenAI's GPT-6 Astra (announced 2026-09-03 per third-party coverage; figures as reported by OpenAI). Anthropic warns benchmark margins are a less reliable guide than real use, where the Opus 5.5 vs Fable 5.1 gap is narrower. Sonnet 5.5 and Haiku 5.5 were promised 'in the coming weeks'.
Sources
- www.anthropic.com/news/claude-opus-5-5
- platform.claude.com/docs/en/models/overview
- platform.claude.com/docs/en/release-notes/overview
This record was checked against its sources on 6 October 2026. How we check