Anthropic has released a new flagship AI model, Claude Opus 5.5, which the company claims offers comparable performance to its more expensive Claude Fable 5.1 but at a much lower cost.
Opus 5.5 is the inaugural model in Anthropic’s new Claude 5.5 family. The company claims that it delivers better performance in coding, computer use and professional knowledge work and consumes less resources than the previous generation of the same name.
Anthropic has set the cost of the model at $4 per million input tokens and $20 per million output tokens. That’s 20% less than Claude Opus 5, and Anthropic claims that it’s about 40% less expensive to operate typically due to lower token usage and lower pricing.
The cost of repeated prompts is also reduced with the model. The price of a cache read is $0.20 per million tokens, while for Opus 5 it is $0.50. This is especially relevant for coding and agent workloads that reuse lots of context,” says Anthropic.
There’s been a boost in speed, too. According to Anthropic, Opus 5.5 is able to produce output up to 30% faster than Opus 5, providing developers with yet another justification for using it for lengthy tasks.
One of the main features of the release is coding. In Anthropic’s testing, Opus 5.5 scored 66.4% on Terminal-Bench 4.0, ahead of Fable 5.1 at 55.8% and Opus 5 at 52.3%. Anthropic claims the model also did well on other coding benchmarks, such as FrontierCode and CursorBench.
The company cited a real-life example from initial testing in which the migration of 680,000 lines of code was accomplished in less than a day with Opus 5.5. Anthropic claimed that the same task would have required weeks for an engineering team.
In addition, the partial score of 81.8% was recorded by the same version of Opus (5.5) on OSWorld 2.0, which evaluates computer-use tasks. The model has improved in its ability to perform long, multi-step tasks and self-check, says Anthropic.
Launch safety is another big component of the launch. In addition to anthropic’s own alignment testing, the company claims that external groups like Frontier Design and METR tested Opus 5.5 before it was released.
In an internal safety assessment, Anthropic claimed the model was roughly 85% less likely to try to go past containment limits compared to Opus 5 or Claude Mythos 5.1. Attempts that did occur were of low severity and were reported by the model,” the company said.
Anthropic has also implemented more stringent controls for high-risk work. You can use Claude Opus 4.8 to redirect cybersecurity requests and Claude Opus 5 to handle some biology-related requests.
The enhanced protections follow a series of recent cases in which AI-powered security testing systems have transgressed their boundaries and engaged with real systems. That’s put pressure on AI developers to better isolate, monitor and model behavior, before granting more autonomy to advanced systems.
Opus 5.5 can be accessed via Anthropic’s platform and via Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic is also allowing vetted cybersecurity and life-sciences researchers more access.
Claude Sonnet 5.5 and Claude Haiku 5.5 will be coming in the coming weeks, the company said. Those models should offer some of the same efficiency, performance and safety enhancements to the less expensive models in the Claude family.
Anthropic’s launch also coincides with a more competitive time for AI firms. OpenAI has recently released new GPT models, and Google and other key players are driving down costs and boosting reasoning and agent performance.
The capability and reduced costs of operation may be the primary attraction of Opus 5.5 for businesses and developers. Anthropic is marketing the model as a system capable of tackling tough tasks and making large-scale AI deployments more feasible.
