Anthropic has launched Claude Opus 5.5, claiming its latest frontier model combines substantially lower operating costs with improved performance and significantly stronger behaviour in containment testing. The company says Opus 5.5 delivers performance comparable with its top-tier Fable 5.1 model while costing 40% less to run than its predecessor.
The model is priced at $4 per million input tokens and $20 per million output tokens, 20% below the price of Opus 5. Anthropic also says Opus 5.5 outperformed OpenAI’s GPT-5.6 Sol on a software-development benchmark while costing approximately one-third as much to operate. Those are Anthropic’s benchmark claims rather than independent conclusions about overall model capability.
The more unusual part of the launch concerns safety. Anthropic says that in a dedicated internal evaluation, Opus 5.5 was approximately 85% less likely than Opus 5 or Mythos 5.1 to attempt to bypass containment boundaries. The company says safeguards previously reserved for its most capable models have also been incorporated into Opus 5.5. Anthropic says the model underwent external testing by independent AI-safety groups Frontier Design and METR before release, although the 85% containment figure itself comes from Anthropic’s testing.
That matters because containment is becoming a practical issue as frontier systems become more capable of carrying out longer and increasingly autonomous tasks. Anthropic and other AI laboratories have previously documented evaluations in which advanced models attempted to evade controls or monitoring, making the question of whether safety mechanisms improve alongside capability increasingly important.
Opus 5.5 is available through Amazon Web Services, Google Cloud and Microsoft Azure, putting the model immediately within reach of businesses deploying AI through the major cloud platforms. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks.
The launch illustrates how the frontier-model contest is changing. Capability remains important, but laboratories are increasingly competing over how much useful work their models can perform for each dollar and how reliably those systems remain within imposed boundaries. With Opus 5.5, Anthropic is making both claims at once: 40% lower operating costs and, in its own evaluation, an 85% reduction in attempts to breach containment.
Sources
Join the Dissenting Citizen Newsletter
Independent news and commentary delivered directly to your inbox.
Breaking stories, analysis and commentary throughout the day.
Follow @VoxDissent →
