Anthropic has launched Claude Opus 5, its new flagship model for complex coding, research and agentic work. The model is available through Claude, the Claude API and major cloud channels. Anthropic says the release improves the model’s ability to plan, use tools, verify its own work and stay effective across long-running tasks.
The practical significance is less about a single benchmark and more about how the model behaves during sustained work. Anthropic is pitching Opus 5 as a system that can take a broad objective, break it into steps, inspect the result and keep iterating. That pattern matters for software projects, analytical work and autonomous agents, where a plausible first answer is not enough.
What changes with Opus 5
Anthropic says Opus 5 approaches the capability of its larger Fable 5 research model at roughly half the price. The company has kept standard API pricing at US$5 per million input tokens and US$25 per million output tokens, matching Claude Opus 4.8. That gives existing Opus customers a relatively direct upgrade path, although total cost will still depend on prompt size, tool use and how long an agent runs.
A separate fast mode is available through the Claude platform and credits. Anthropic says it can deliver about 2.5 times the generation speed at twice the base token price. This creates a clear latency-versus-cost choice: interactive coding and time-sensitive analysis may justify fast mode, while background jobs may be better suited to the standard service.
On Claude’s consumer plans, Opus 5 becomes the default model for Max subscribers and the strongest model available to Pro subscribers. API users can select the model with the identifier claude-opus-5. Organisations should still test the change against their own prompts, tools and evaluation sets before replacing a production model.
Designed for agents and coding workflows
The launch focuses heavily on software engineering and knowledge work. Anthropic says Opus 5 is better at navigating large codebases, making coordinated changes and checking whether those changes work. It also highlights gains in scientific reasoning, visual output and tasks that require information to be gathered from several tools.
Two beta features are intended to reduce friction in longer sessions. Developers can change the tools available during a conversation without automatically invalidating the prompt cache, and automatic safety fallbacks can help an application recover when a requested action is blocked. Those controls could make multi-step agents less brittle, but teams will need to observe how fallback behaviour affects accuracy and auditability.
Anthropic’s claims should be treated as vendor-reported results until independently reproduced. Aggregate benchmark scores also do not answer important operational questions such as latency under load, error recovery, regional availability and performance on private domain data.
Availability beyond Anthropic
AWS has separately confirmed that Claude Opus 5 is available through Amazon Bedrock. AWS says the model can be used with Bedrock features including Guardrails and Knowledge Bases, and that zero data retention applies by default. Buyers should confirm the exact regions, residency options and commercial terms that apply to their account.
Legal AI provider Harvey has also made Opus 5 available in its platform. Harvey reports that the model produced similar results to an Opus 4.8 configuration while using less reasoning effort and fewer tokens in its internal testing. The rollout starts in the United States, with European and Australian availability expected later. These are supplier results rather than a general guarantee of performance.
Safety and deployment considerations
Anthropic says its testing did not find a new frontier-level increase in risky capabilities and that Opus 5 remains behind Mythos 5 for offensive cyber tasks. The company nevertheless applies safeguards and provides a confidential-computing option for workloads that need stronger isolation. Customers should review Anthropic’s system documentation alongside their own access controls, logging and human-approval policies.
For teams already using Claude, the most useful next step is a controlled comparison with the current production model. Evaluations should include task completion, tool-call accuracy, token consumption, recovery after errors and the quality of the final checked result. A model that performs better on the first response may not be the best choice if it consumes substantially more resources during a long agent run.
Opus 5 is therefore best understood as an upgrade aimed at sustained, verifiable work. Anthropic has kept the headline token price stable and added options for greater speed, but the real value will depend on whether the model can complete more of a workflow with fewer interventions.