On September 22, 2026, Anthropic officially announced Claude Opus 5.5 through its Newsroom — the first model in the new 5.5 family. According to the company, it performs at the level of the top-tier Claude Fable 5.1 on most work, while costing 40% less to run on typical workloads than its predecessor Opus 5, generating output more than 30% faster, and cutting API pricing from $5/$25 to $4/$20 per million input/output tokens — a 20% reduction.
Stronger, not pricier: the selling point is efficiency
This is not a routine "bigger and stronger" upgrade. Anthropic's comparison table shows Opus 5.5 leading on multiple agentic coding and knowledge-work benchmarks — Terminal-Bench 4.0 (66.4%), FrontierCode v1.1, and GDPval-AA v2.1 — including against OpenAI's GPT-6 Astra. Independent evaluator Artificial Analysis scored it 58 on its Intelligence Index — the highest score ever recorded on that index, about five points ahead of Fable 5.1 and GPT-6 Astra.
The price numbers deserve an asterisk, though. Artificial Analysis also found that at its maximum setting, Opus 5.5 burns roughly 119,000 tokens per benchmark task, far more than GPT-6 Astra's 27,000. Token prices fell 20%, but the model is more liberal with tokens, so per-task cost at max settings is roughly in line with Opus 5. The real savings come at medium effort and below — Artificial Analysis concluded that four of Opus 5.5's five effort levels sit on the intelligence-vs-cost Pareto frontier: among models scoring above 50, it is either cheaper or stronger.
A safety report published alongside the model, for once
Just as notable as the pricing is how prominently Anthropic put safety evaluation on display. Before release, Opus 5.5 was tested by two external organizations, METR and Frontier Design, and achieved the highest score ever in Anthropic's internal automated behavioral audit. Testing by security firm Gray Swan showed it ties Fable 5.1 for the lowest prompt-injection success rate of any model tested. The accompanying system card discloses the evaluation in full, including the finding that its biology and cybersecurity capabilities approach Fable 5.1 — which is why it ships with equivalent safeguards: vetted research organizations can apply to the Life Sciences Verification Program to use it for biology research.
The takeaway: the first answer after the slow-down manifesto
This is the first major release since Anthropic CEO Dario Amodei publicly called for the industry to slow the pace of frontier AI development. The posture of this answer is clear: match the flagship on capability, push prices down, and lead with safety transparency. The new version of Claude Code already sets it as the default Opus model, five-hour usage limits for Pro, Max, Team, and seat-based Enterprise plans have been raised, and Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks.
For the industry, the signal is direct: frontier-model competition is shifting from pure benchmark scores to the cost of completing a real task. When the strongest model also gets cheaper and more transparent, developers will have to redo their math.
