Anthropic launches Claude Opus 5 with frontier-level performance at lower cost

Anthropic has released Claude Opus 5, a new model positioned close to the frontier capabilities of Claude Fable 5 while costing about half as much. The company says it is available immediately and is now the default model for Claude Max subscribers, as well as the strongest option on Claude Pro.

Built for AI workflows

Anthropic reports state-of-the-art results for Opus 5 on coding and knowledge-work evaluations including Frontier-Bench and GDPval-AA, although the model trails Mythos 5 on cybersecurity tasks. The company highlights particularly strong cost-performance on software engineering: Opus 5 more than doubles Opus 4.8’s Frontier-Bench result and performs within 0.5% of Fable 5’s peak CursorBench score at maximum effort, for half the cost per task.

The model also targets automation and computer-use applications. Anthropic says it leads on ARC-AGI 3, Zapier AutomationBench, and OSWorld 2.0 at comparable costs. Its configurable effort settings let developers trade speed and token usage for more intensive reasoning.

Opus 5 is also reported to improve on life-science tasks, including organic chemistry, structural biology, bioinformatics, and protein-function prediction. Anthropic demonstrates the model producing interactive visual artifacts such as a wind-tunnel simulation and a 3D animal-cell illustration.

More autonomous problem-solving

Early examples emphasize verification and persistence. Opus 5 reportedly reconstructed a 3D machine part after creating its own computer-vision pipeline, identified the root cause of a package-manager bug, and built a test harness while developing a market-data feed without a live reference feed.

For developers, the release suggests a shift toward models that can manage longer, less structured tasks—not just generate code, but test, inspect, debug, and iterate on their work.

Source: Anthropic News