Anthropic has launched Claude Opus 5, now available across all of its platforms. It is the new default model for Claude Max and the strongest model on Claude Pro. Developers can access claude-opus-5 through the Claude API. Pricing remains at $5 per million input tokens and $25 per million output tokens, matching Opus 4.8, while fast mode runs at about 2.5 times the default speed for twice the base price.
Built for daily coding and knowledge work, Opus 5 is designed to approach Claude Fable 5's intelligence at half the price. Its effort setting lets customers trade token use and cost against speed and capability. Anthropic says the model leads Frontier-Bench v0.1 and more than doubles Opus 4.8's result at a lower cost per task. At maximum effort on CursorBench 3.2, it finished within 0.5% of Fable 5's peak score for half the cost per task.
Anthropic also reports that Opus 5 scored three times as high as the next-best model on ARC-AGI 3 and beat every other model at any given cost on OSWorld 2.0. It surpassed Opus 4.8 across Anthropic's life-sciences evaluations and produced stronger visual output. The model places more emphasis on checking its work: in one example, it built a computer-vision pipeline to reconstruct a machine-part drawing as a 3D FreeCAD model after repeated attempts.
Early-access customers reported strengths in difficult debugging, analytics, financial research, genomics, legal agent work, code review, and long-running coding tasks. Anthropic says Opus 5 is its most aligned model so far and does not move the frontier in risky dual-use capabilities. It remains behind Mythos 5 in biology research and offensive cybersecurity.
For Anthropic, the release pairs wider access with narrower cyber controls. Classifiers permit vulnerability discovery in source code but block binary scanning, penetration testing, and exploit generation; flagged requests fall back to Opus 4.8 by default in Claude AI, Claude Code, and Claude Cowork. Two beta updates also let API developers change tools mid-conversation without invalidating the prompt cache and route safety-flagged requests to another model.