On September 22, 2026, Anthropic released Claude Opus 5.5, the first model in the new Claude 5.5 lineup. According to the company, it performs at the level of the flagship Claude Fable 5.1 on most tasks, while being 40% cheaper than Opus 5 for typical token-metered workloads. Pricing has already been published, and the model is immediately available on Pro, Max, Team, and Enterprise plans, via the Claude API, as well as on AWS, Google Cloud, and Microsoft Foundry.

image
image

What Happened

Anthropic introduced Claude Opus 5.5 and published its terms alongside the announcement: $4 per million input tokens and $20 per million output tokens, which is 20% lower than Opus 5, with cache reads at $0.20 per million, down 60%. In Claude Code and on the Claude Platform, a fast mode is available with a claimed speedup of up to 2.5x at a price of $8 and $40 per million input and output tokens, respectively. In the announcement's own benchmarks, the model scores 66.4% on Terminal-Bench 4.0, compared to 55.8% for Fable 5.1 and 52.3% for Opus 5. A system card describing the included production safeguards was also published, in which Anthropic itself warns: at this level of models, metric gaps are unreliable, and the actual gap with Fable 5.1 is narrower than the numbers suggest.

Context

The key to understanding the release is the economics of agentic scenarios. In long, multi-step processes, a model takes hundreds of steps and repeatedly rereads the same context, so cache read costs often determine the final bill. A 20% reduction in token price and a 60% reduction in cache read costs, combined with fewer tokens per task, yield the claimed 40% savings for typical workloads. For Anthropic, this is the first significant price reduction in the flagship Opus lineup. The publication of a system card with caveats about production safeguards and the launch of the 5.5 branch, to be followed by Sonnet 5.5 and Haiku 5.5, indicate a shift toward more frequent point releases, where the emphasis is on launch cost rather than just model quality.

Why This Matters for the Industry

For companies using agentic scenarios, the unit economics of the frontier have improved: flagship-level performance at token prices 20% lower than Opus 5, cache reads 60% cheaper, and in Anthropic's published case studies, Optiver, Kiro/AWS, and GitHub Copilot report the same results with 40–50% lower costs and half the number of steps. This directly reduces the cost of long, multi-step processes where inference expenses dominate. After Sonnet 5.5 and Haiku 5.5 are released in the coming weeks, the lineup will cover the entire price range, and routing between models by task step — cheap steps on Haiku, complex ones on Opus 5.5 — will become a standard cost-optimization pattern. A price response from competitors is likely, and value will shift to the product layer above the API: workflow design, memory, and orchestration.

Why This Matters for Users

If you use Claude Code, the Claude API, or Pro, Max, Team, or Enterprise subscriptions, Opus 5.5 is already available — you can run it on your tasks and compare it with Opus 5. A practical step this week: run 2–3 key scenarios on both models and measure three numbers — task cost, number of steps, and cache read share. The greatest gains are expected for cache-heavy agentic processes and interactive code work via fast mode in Claude Code. Start by choosing non-critical workflows and recording cost per task before and after switching.

What Is Still Unknown / Limitations

Parity with Fable 5.1 and the speedup in fast mode are vendor measurements without independent verification: the available materials do not disclose the benchmark methodology, including the number of runs, result variance, and the scaffolding used. The phrasing "at the level of Fable 5.1 on most tasks" is not operationally defined — there is no task distribution or threshold defining "most." Independent runs on third-party task sets will appear later; until then, the Terminal-Bench 4.0 figures should be treated as hypotheses to test on your own traffic, not as established results.

Sources

Author

Look at AI, editorial team