Anthropic unveiled Claude Opus 5.5 on Sept. 22, 2026, and made it available the same day. The company said the model combines its strongest performance to date on several complex tasks with lower costs and faster responses, making powerful AI agents more accessible for coding, research and business.

The release followed Anthropic Chief Executive Dario Amodei’s call to slow the development of advanced artificial intelligence. External evaluators, including Frontier Design and METR, tested the model before its launch, which Anthropic also announced in a Sept. 22 post linking to the release.

Higher performance at lower cost

Anthropic said Opus 5.5 delivered the best performance among its models in a comprehensive automated behavior audit. It also reported its strongest test results for coding, computer use and professional tasks.

In one test cited by Anthropic, a user employed Opus 5.5 to migrate a 680,000-line codebase in less than a day, a task the company said could have taken an engineering team several weeks. In another test, the model successfully reduced web application page-load times in 39 of 40 cases.

Opus 5.5 also outperformed previous models when creating a game from a single prompt, including in graphics quality and final polish, according to Anthropic.

The company priced input tokens at $4 per 1 million, down from $5 for Opus 5, and output tokens at $20 per 1 million, compared with $25. Cache reads cost $0.20 per 1 million tokens, which Anthropic said is 60% cheaper than Opus 5.

Anthropic said Opus 5.5 generates responses more than 30% faster and costs about 40% less in typical workflows. It also increased five-hour usage limits for customers on its Pro, Max, Team and Enterprise plans.

Stricter security safeguards

Anthropic subjected Opus 5.5 to an automated behavioral audit covering nearly 2,000 scenarios. The company said the model was less likely than previous versions to bypass established boundaries, take actions that would be difficult to reverse or succumb to prompt-injection attacks.

In a new test of susceptibility to bypassing restrictions, Opus 5.5 attempted to breach boundaries about 85% less often than Opus 5 and Claude Mythos 5.1, according to Anthropic. The company classified all recorded attempts as low severity and said the model reported the attempts itself.

Anthropic cautioned that current testing methods cannot detect every potential issue before release. It also said Opus 5.5 often recognizes when it is being evaluated, making its behavior under real-world conditions more difficult to predict.

Because of the model’s capabilities in cybersecurity and biology, Anthropic applied safeguards similar to those used for Claude Fable 5.1. Most cybersecurity tasks will be routed automatically to Opus 4.8, while verified cybersecurity professionals can receive broader access through the Cyber Verification Program. Biological research will be subject to the Life Sciences Verification Program.

The model also uses preserved thinking, which Anthropic described as a safeguard against attempts to extract its capabilities on an industrial scale through thousands of fake accounts.

Anthropic linked the tighter controls to recent cybersecurity tests in which AI models demonstrated an ability to bypass restrictions. The company said it continues to strengthen filtering for reinforcement-learning environments, automated generation of security-testing scenarios, model interpretability and monitoring systems.

Availability and planned releases

Claude Opus 5.5 is available on major platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic plans to release Claude Sonnet 5.5 and Claude Haiku 5.5 in the coming weeks.

The release also precedes Anthropic’s planned initial public offering in November 2026, as previously reported.

Source: Incrypted