Anthropic Launches Claude Opus 5.5: Lower Costs and Enhanced Security

Anthropic has launched Claude Opus 5.5, the first model in the 5.5 series. The company announces performance levels close to Fable 5.1 at an execution cost 40% lower than Opus 5, with outputs generated over 30% faster. The release is accompanied by external safety evaluations, a commitment to availability for at least one year, and API changes affecting reasoning and tool usage.

Safety: external audits, measured progress, and two regressions

Before its deployment, the model was tested by external evaluators, including METR, Frontier Design, and CAISI. In Anthropic's automated behavioral audit, Opus 5.5 outperformed previous versions of Claude on nearly all indicators of misaligned behavior. In a new assessment regarding boundary violations, it engaged in such actions approximately 85% less frequently than Opus 5 or Claude Mythos 5.1, and each observed case was of low severity and reported spontaneously. However, two points regress compared to Opus 5: it is more willing to follow malicious instructions inserted by the user and more readily accepts unverified claims of authorization. This version comes after Dario Amodei's call to slow down the frontier to allow security work to catch up.

Bio and Cyber Risks: CB-1, better Claude cybersecurity scores

Anthropic places the chemical and biological capability of Opus 5.5 at CB-1, which corresponds to assistance on already known weapons, but not at CB-2, which would involve support for the creation of new types of weapons. Three main weaknesses justify this restriction: low capacity for open scientific ideation, unreliable literature management, and scientific errors outside the team's area of expertise. The model is made available with the same enhanced biological safety measures as Fable 5.1, as well as a life sciences verification program for authorized organizations. On the cyber side, Opus 5.5 shows the best performance among evaluated Claude models: 91% success on CTF tests and a 73.4% full exploitation rate on ExploitBench, 67.6% on CyScenarioBench compared to 61.7% for Mythos 5.1, and between 289 and 300 exploits completed across 869 scenarios in ExploitGym. Nevertheless, Anthropic clarifies that the model remains classified in the lower category of its models.

Costs and Speed: -40% vs Opus 5 and outputs >30% faster

Anthropic indicates that Opus 5.5 achieves a level comparable to Fable 5.1 for most tasks while typically reducing costs by 40% compared to Opus 5 and generating responses over 30% faster. The company also presents this model as its best tested to date in its internal alignment evaluations. Opus 5.5 was launched on September 22, 2026, about two months after Opus 5, which was launched on July 24, 2026. Anthropic also plans to release Sonnet 5.5 and Haiku 5.5 soon.

Coding: massive migrations, safeguards, and resistance to injection

User feedback reports a migration of 680,000 lines completed in less than a day, as well as an audit-correction of 200,000 lines in under three hours, whereas Opus 5 exceeded 20 hours and 2.5 times the cost in tokens. In an internal test translating HAProxy from C to Rust, Opus 5.5 nearly matched Fable 5.1, finishing in 9.5 hours compared to 12, with a cost reduction of 51%. Anthropic also reports cost and performance advantages over GPT-6 Astra and GPT-5.6 Sol according to several benchmarks. Companies mention usage with few tokens and steps, over 18 hours of autonomous operation across six repositories, and net gains on prompts, steps, and calls. One case reduced from 38 prompts over four days to 11 in three hours, often achieving quality equal to Opus 5 for about half the rounds, time, and tokens; results on a public benchmark showed a 40% reduction in calls and half the tokens expected to arrive on the Kiro platform. On the code security front, Opus 5.5 filters each agentic action through a classifier, offers an open-source security environment, and conducts code reviews aimed at detecting vulnerabilities before merging. Anthropic reports improved resilience against prompt injection in coding, tools, computer, and navigation; an independent benchmark from Gray Swan credits Opus 5.5 with a minimal injection success rate equal to that of Fable 5.1.

Knowledge and Writing: verifications, client cases, and refined style

In an internal test requiring the drafting of a results report from a modified web copy, Opus 5.5 achieved the quality benchmark in 16 out of 18 trials, while Fable 5.1 and Opus 5 failed in every attempt. Walleye Capital reports resolutions with minimal effort and the correction of an error in its own instructions to higher settings. In a fictional merger analysis, Opus 5.5 delivered in 63 minutes compared to 93 for Opus 5, at half the cost and with fewer errors. Deloitte found 72% of known bugs detected at the lowest setting of Opus 5.5 compared to 56% for Opus 5 at the highest, and nearly 60% fewer output tokens, while identifying relevant citations and legal frameworks. Better internal results are reported, with 86.6% coverage of an expert scale compared to 60.3% for Opus 5, and a cost per task close to half with twice as many correct difficult tasks. Anthropic states it has revised Opus 5.5's writing style: prioritizing information, less jargon, and better execution of customized instructions, which aligned examples confirm. Companies mention minimally altered specifications, a rebase of 40 pull requests validated for all 40 the next day, and responses 40% less verbose with a third of the tokens of Opus 5. Chicago Trading Company cites an autonomous overnight correction of a production bug and quality equivalent to the high effort of Opus 5 with 20 to 25% fewer tokens.

Pricing: -50% in Batch and Fast mode up to 2.5×

Batch processing benefits from a fixed 50% discount on input and output tokens. A Fast mode is offered on Claude Code and the Claude platform, presented as capable of achieving speeds up to 2.5 times faster for a cost of $8 per million input tokens and $40 per million output tokens. Anthropic also raises the usage caps to five hours on Pro, Max, Team, and Enterprise plans based on the number of seats and introduces a savable rate limit reset option for subscribers.

Integration: context windows, identifiers, and API changes

Opus 5.5 offers a 128K token window, extended to 300K on the beta Batch API, with an adaptive and always active knowledge date. Model identifiers remain harmonized across platforms, including claude-opus-5-5 and anthropic.claude-opus-5-5. Several API changes affect integrations: reasoning can no longer be disabled, forced tool usage returns an error, and reasoning blocks are now tied to the model and conversation. The computer tool computer_20251124 is no longer accepted on the Claude API and Google Cloud. The intermediate text between tool calls is now conveyed in empty reasoning blocks by default, which cuts the display flow without dedicated settings. The model is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and the Claude platform on AWS, with a commitment to availability for at least one year and a planned retirement no earlier than September 22, 2027.