Anthropic has launched Opus 5, the latest version of the model that has become a go-to option for coding and other software development work.

Benchmarks such as Frontier-Bench and DeepSWE place Opus 5 at roughly the same level as, or slightly ahead of, Anthropic's Fable model on coding tasks, and ahead of Opus 4.8 and OpenAI's GPT-5.6-Sol in most categories. The improvements are iterative rather than a leap on the scale of Opus 4.5; the main selling point is delivering performance just under Fable at approximately half the cost.

Anthropic deliberately withheld cutting-edge cybersecurity training from Opus 5. The company says the model is reasonably good at spotting vulnerabilities but is "substantially behind Mythos 5" at exploiting them, and it does not carry Fable's controversial safeguards such as 30-day data retention for incident review.

Pricing is $5 per million input tokens and $25 per million output tokens — the same as its predecessor and cheaper than Fable. Competitive pressure is real, though: the Chinese open-weight model Kimi K3 offers similar performance at $15 per million output tokens.

The release reflects a broader industry shift toward cost efficiency. Companies including Cursor and Meta are building "model routers" that pick smaller or cheaper models per prompt, and Anthropic faces pressure to keep lowering token costs or risk losing routine development workloads to open-weight and local alternatives.