Introducing Claude Opus 5

Anthropic has released Claude Opus 5, its newest model designed for high-level coding and complex knowledge work. The model is available starting today and provides significantly improved performance over the previous Opus 4.8 version at the same cost. According to internal benchmarks, Opus 5 leads on evaluations like Frontier-Bench and performs well on computer use tasks.

Engineers can expect the model to verify its own work and iterate through difficult problems with greater agency. Early users report that Opus 5 handles long-horizon tasks better than its predecessors, catching logical errors during the planning phase. It is now the default model for Claude Max and the strongest option for Claude Pro subscribers.

From a safety perspective, the company describes this as its most aligned model yet. It adheres closely to its constitutional guidelines and remains limited in its ability to generate harmful exploits. Anthropic has maintained specific safeguards that allow the model to assist with legitimate security research while preventing misuse.

Developers can access the model via the API now, with pricing set at $5 per million input tokens and $25 per million output tokens. The release also includes new features for the platform, such as mid-conversation tool changes and automatic API fallbacks to ensure continuity during complex tasks.