Anthropic released Claude Opus 5.5 on September 22, 2026, introducing a combination that is uncommon in AI model launches: the new system costs less to operate, generates output faster, and is more resistant to manipulation than the Claude Opus 5 model it replaces, which shipped only two months earlier. The company put the headline figure at a 40% reduction in overall operating costs.

What Changed and What It Costs Now

The price cuts are most visible in Anthropic's updated rate card. Input tokens are now priced at $4 per million and output tokens at $20 per million, each representing a 20% reduction from Opus 5's rates of $5 and $25. Cache reads fell even further, dropping 60% to $0.20 per million tokens. For multi-step agenticloads, where the same context gets read repeatedly across chained calls, that cache discount is where per-task savings compound.

Output generation is more than 30% faster than the previous model. That difference carries limited weight for casual queries but becomes significant in agentic workflows that chain dozens of model calls together. Anthropic built Opus 5.5 specifically for long-duration agentic coding and professional workloads, equipping it with a 1M-token context window and a maximum synchronous output size of 128K tokens.

On benchmarks, the model matches or exceeds Claude Fable 5.1, Anthropic's higher-tier offering, across a range of coding and knowledge-work evaluations. On GDPval-AA v2.1, it scores 1,846 Elo — effectively flagship-tier results at a lower price point than the model it replaced.

Safety Improvements That Are Actually Measurable

Independent evaluators, including Frontier Design and METR, found that boundary-circumvention attempts fell 85% compared with both Opus 5 and Mythos 5.1. Those are external assessments rather than self-reported figures, a distinction that typically carries more weight with buyers and researchers. Anthropic is billing Opus 5.5 as its strongest-performing model in comprehensive alignment testing to date.

CEO Dario Amodei has publicly pushed for a more measured pace of capability releases given ongoing safety concerns across the industry, and the emphasis on alignment in this launch reflects that position. The two-month release gap makes that tension concrete: rapid shipping on one side, external safety validation on the other. Anthropic also says the model now prioritizes essential information and reduces jargon.

Who Gets Access and Through What Channels

Opus 5.5 is available immediately through the Claude API, the Claude Platform, Amazon Bedrock, and Google Cloud. Microsoft is also listed as a distribution partner, giving the model coverage across the three dominant enterprise cloud environments simultaneously at launch — meaning organizations standardized on a single provider can adopt it without adding a second vendor relationship.

Anthropic expanded usage limits alongside the release. Pro, Max, Team, and Enterprise subscribers receive increased five-hour usage caps, with optional rate-limit resets available — headroom that aligns with the model's positioning for long-duration workloads.

Early user reports highlight code migration tasks as a standout use case, with large-scale refactoring and framework migrations completing with fewer interruptions — consistent with the model's design intent for professional developer tooling. With pricing, benchmark, and safety figures now on the record, broader production use across the expanded subscriber base will supply the wider data against which those launch-day numbers are judged.