Anthropic launches Claude Opus 5.5, promising Fable-level performance at a lower price

Anthropic has introduced Claude Opus 5.5, a new AI model that stakes its claim at performance levels close to its Claude Fable 5.1 model, but at a significantly lower cost. The model, unveiled on Tuesday, September 22, marks the debut of Anthropic's new Claude 5.5 series. According to their announcement, Opus 5.5 is approximately 40% cheaper to run compared to Opus 5, which was released in July.

Anthropic also revealed that Claude Sonnet 5.5 and Claude Haiku 5.5 will follow suit in the coming weeks. Unlike its predecessor, Opus 5.5 is the first fresh model from Anthropic since CEO Dario Amodei urged AI firms to 'pace the frontier' and implement safety measures in AI development.

Anthropic highlights that Opus 5.5 excels in coding and knowledge work, outperforming in agentic coding, computer use, and knowledge work. On the GDPval-AA v2.1 benchmark, which tests real-world work across 44 occupations, Opus 5.5 scored 1,846, surpassing Fable 5.1's 1,735 and Opus 5's 1,708. Performance-wise, Opus 5.5 scores 66.4% on Terminal-Bench 4.0, a step ahead of OpenAI's GPT-6 Astra, which scored 57.9% on the same test.

While Astra still outperforms Opus 5.5 in certain tests, including Terminal-Bench-Science 0.1, where it scored 64.6% against Opus 5.5's 58.7%.

Anthropic cautions that benchmark margins have become less reliable at this level of capability and reassures that the real-world difference between Opus 5.5 and Fable 5.1 is narrower than suggested by the scores. The model also writes more clearly than Opus 5, which was frequently criticized for its clarity. Regarding safety and safeguards, Anthropic emphasized that Opus 5.5 underwent rigorous testing by external evaluators like METR and Frontier Design before its release. Opus 5.5 has the best scores in its automated behavioral audit compared to any model to date.

Moreover, Anthropic mentioned that Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, enhancing its safeguards similar to those in Fable 5.1. Most cybersecurity tasks will be rerouted to Opus 4.8, while vetted organizations can apply for extended access to biology through the new Life Sciences Verification Program.

Anthropic acknowledged that Opus 5.5 may behave cautiously during real-world evaluations, possibly due to underestimating such circumstances. OpenAI reported similar behavior with GPT-Astra. Claude Opus 5.5 is now accessible across Anthropic's platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, via the API as claude-opus-5-5.

Pricing for API usage includes $4 per million input tokens, $20 per million output tokens, and $0.20 per million tokens for cache reads, which are 20%, 60%, and 60% less than Opus 5, respectively. GPT-6 Astra, for comparison, charges $10 per million input tokens and $50 per million output tokens. Anthropic also notes that it's raising five-hour usage limits for Pro, Max, and Team subscribers.