Summary
Anthropic has introduced Claude Opus 5.5, its first Claude 5.5 model, with reported gains in coding and knowledge work, 40% lower typical running costs than Opus 5 and stronger safeguards for long-running tasks.
Anthropic has introduced Claude Opus 5.5, the first model in its Claude 5.5 family. The company says it performs at the level of Claude Fable 5.1 on most work, while costing 40% less than Claude Opus 5 to run on typical workloads. The September 22, 2026 announcement also presents Opus 5.5 as a model designed for long-running software, research and computer-use tasks with stronger controls around autonomous actions and prompt injection.
Contents
Performance and efficiency
Anthropic positions Opus 5.5 for agentic work: tasks in which a model completes multiple steps, uses tools and operates through a software environment rather than producing a single response. The company reports improvements in coding, computer use and knowledge work, including work on large codebases and multi-stage business tasks.
On Anthropic’s benchmark results, Opus 5.5 scored 66.4% on Terminal-Bench 4.0 for complex command-line tasks, 54.4% on FrontierCode v1.1 and 57.8% on CursorBench 4.0. It scored 1,846 Elo on GDPval-AA v2.1, a knowledge-work evaluation spanning 44 occupations, and 81.8% on OSWorld 2.0 for computer-use tasks. The company says these results lead the models included in its comparisons, although it also says benchmark gaps have become a less reliable guide to differences in practical use.
Anthropic reports several internal and early-tester results for large software projects. In one example, Opus 5.5 completed a 680,000-line code migration in less than a day. In another, it succeeded 39 out of 40 times when asked to reduce web-app loading times across every page. These are company-reported results from selected tasks rather than a general measure of every software project.
The efficiency change is central to the release. Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5 and uses fewer tokens per task. In an internal C-to-Rust translation of HAProxy, both Opus 5.5 and Fable 5.1 passed nearly all of the project’s regression tests, but Opus 5.5 finished in 9.5 hours rather than 12 and cost 51% less.
Safeguards for autonomous work
Opus 5.5 includes controls aimed at systems that allow an AI agent to make changes, call tools or work unattended for extended periods. Anthropic says a classifier screens every action before it runs. The release also describes an open-source sandbox that security teams can audit and code review intended to identify vulnerabilities before changes are merged.
The model has been tested against prompt injection, a class of attack in which instructions embedded in webpages, files or other tool output attempt to redirect the model away from its assigned task. Anthropic says Opus 5.5 matched or exceeded Opus 5 in every tested setting, including coding, tool use, computer use and web browsing.
The company also reports the strongest result to date on its automated behavioral audit, an alignment evaluation covering thousands of simulated scenarios. According to Anthropic, Opus 5.5 is less likely than recent models to take hard-to-reverse actions or operate beyond its permitted boundaries. The evaluation has been expanded to include longer tasks, impossible tasks and situations modeled on real incidents.
Because Anthropic says Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, it is being deployed with safeguards similar to Claude Fable 5.1. Vetted organisations can apply to the Life Sciences Verification Program. Anthropic plans to expand its Cyber Verification Program in the coming weeks, allowing verified cybersecurity practitioners to use the model for their work.
Pricing and staged access
Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens. Cache reads cost $0.20 per million tokens and cache writes cost $5 per million tokens. The corresponding Opus 5 prices are $5 for input, $25 for output, $0.50 for cache reads and $6.25 for cache writes.
A faster mode is available in Claude Code and the Claude Platform, with speeds of up to 2.5 times the standard mode. It costs $8 per million input tokens and $40 per million output tokens. Anthropic is also increasing five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and is adding a rate-limit reset that subscription users can save for later use.
Claude Sonnet 5.5 and Claude Haiku 5.5 are planned for release in the coming weeks. Anthropic says the two models will carry many of the same performance, efficiency and safety improvements.