Anthropic has introduced Claude Opus 5.5 – the first model of the new Claude 5.5 family. Developers emphasize tasks where AI must independently perform complex work for an extended period. According to the company, Opus 5.5 is comparable to Claude Fable 5.1 in most tasks but uses approximately 40% fewer computational resources than Opus 5. The company highlights that the model was released following CEO Anthropic Dario Amodei's call for an approach where the development of advanced AI systems should keep pace with control measures and risk assessments.
The model has been improved for migrating large codebases, auditing programs, optimizing applications, and preparing analytical materials. In one internal test, the system helped transfer 680,000 lines of code in less than a day; Anthropic estimates that a similar task by an engineering team could take several weeks. In the coming weeks, the company plans to release Sonnet 5.5 and Haiku 5.5 with performance, efficiency, and security updates.
Anthropic claims that savings are achieved through lower token prices and a reduction in their total number per task. In typical scenarios, this results in about 40% cost reduction, and the generation speed exceeds Opus 5 by more than 30%. For Pro, Max, and Team plans, usage limits have been increased by five hours, and the option to retain the available reset limit has been added. Among the capabilities is agent programming: in one test, Opus 5.5 improved page loading speed in 39 out of 40 cases. In another experiment, the model outperformed other systems in the quality of graphics and game polish created from a single prompt. The company also reports high results in financial analysis, business reporting, and data research, noting that as modern models' metrics converge, benchmarks increasingly fail to reflect differences in real-world performance.
Anthropic has strengthened the protective mechanisms of Opus 5.5. According to the company, the model showed the best results among tested systems in an internal automated behavior audit, which included thousands of simulated scenarios. Developers note less frequent hard-to-revert actions, increased resistance to prompt injection, more accurate adherence to set boundaries, and improvements in honesty and instruction following. For cybersecurity and biology, restrictions comparable to those used for the company's most powerful models are applied: most cyber requests are redirected to other models, access to biological scenarios is granted to verified organizations through the Life Sciences Verification Program, and the Cyber Verification Program is expanded for cybersecurity specialists. The model also received "preserved thinking" protection, limiting attempts to extract internal reasoning through the API.
Related articles
