Anthropic has released Claude Opus 5.5 with a combination that would have been unusual a year ago: higher capability, lower operating cost and stronger safety controls. The model is the first release in the company’s Claude 5.5 family and is positioned for coding, long-running agents and professional knowledge work.
Anthropic says Opus 5.5 performs at the level of its Fable 5.1 model on most work while costing about 40 percent less to run than Opus 5. Its standard API price is $4 per million input tokens and $20 per million output tokens, with cache reads priced at $0.20 per million tokens.
The pricing change is as important as the model upgrade. AI agents can make dozens or hundreds of model calls during a single job, so the economics of each call quickly become part of a company’s operating budget. Lower token use and lower prices can change which workflows are practical to automate.
Anthropic has also made safety a visible part of the release. Opus 5.5 launches with safeguards covering cybersecurity, biology and model-distillation risks. The company says external groups including Frontier Design and METR evaluated the model before release.
The approach reflects a wider shift in the AI industry. More capable models can interact with computers, write software, inspect systems and carry out multi-step tasks. That makes capability useful, but it also means the failure modes are different from those of a conventional chatbot.
Anthropic says Opus 5.5 is available across Claude’s consumer and business plans and through major cloud platforms including Amazon Web Services, Google Cloud and Microsoft Azure. Developers can also access it directly through the Claude API.
The competitive picture is now moving quickly. OpenAI has released GPT-6 Sol and Luna at lower price points, while Google continues to invest in its own models and custom accelerators. For Anthropic, Opus 5.5 is an attempt to make a frontier-class system economical enough to become a routine production tool rather than an expensive specialist.
The technology is moving quickly, but the commercial test remains familiar. Products have to work consistently, fit existing systems and justify their cost.
The release also puts pressure on the definition of a frontier model. If a newer system can perform comparable work with fewer tokens and lower prices, the competitive advantage is no longer just intelligence in isolation. It becomes efficiency across an entire task. That is particularly relevant to coding agents, where a model can spend thousands of tokens examining files, running tests and correcting its own mistakes.
In production environments, the combination of capability and efficiency can be more important than a headline score.
Anthropic’s safety position is becoming part of its product differentiation. The company is not presenting safeguards as a separate research project that sits beside the model. They are part of the release itself, particularly for domains where an autonomous system could create direct physical or security consequences. That approach will be watched closely as customers give models more authority to act.
For customers, the practical test will be whether the additional safeguards arrive without making the model materially harder to use.
The safety controls also show how model capability is becoming inseparable from deployment policy. Anthropic says Opus 5.5 can hand sensitive cybersecurity and biology work to safer fallback models when its safeguards intervene. That approach accepts that a single model does not need to be responsible for every task. In production systems, routing and permission layers may become as important as the underlying model because they determine which capabilities are available in a particular situation.
The release also shows how AI companies are treating safety as part of production engineering. The relevant question is not whether a model can be made perfectly safe, but whether the system can limit dangerous capabilities when they are not required. Routing sensitive work to a safer model, preserving auditability and controlling access can reduce the consequences of an unexpected model response. Those controls will become more important as agents gain more authority.