Anthropic launches Opus 5: a cheaper, less restricted model that rivals Fable 5 on key benchmarks

Anthropic today released Claude Opus 5, the latest iteration of its long-standing flagship model, just two months after Opus 4.8. The new model arrives in a crowded field: Anthropic rolled out Mythos 5, Fable 5, and Sonnet 5 in June, leaving only Haiku waiting for a 5-series upgrade.

Opus 5 is positioned as a capable, more affordable alternative to Fable 5. While smaller, it actually outperforms Fable 5 on several benchmarks, according to the company. Per-million-token pricing remains unchanged from Opus 4.8 ($5 input / $25 output) — half the price of Fable 5 and slightly cheaper than OpenAI's GPT-5.6. A new "Fast mode," available in research preview, doubles speed for twice the price.

A key selling point is Opus 5's lighter safety touch. Anthropic expects its safety classifiers to engage 85% less often than on Fable 5, reflecting the model's lower capability tier. However, following recent government scrutiny over Fable 5's cyber risks, Opus 5 includes stronger cybersecurity safeguards than Opus 4.8 — the company calls it "the most aligned Opus model." It blocks activities like scanning for vulnerabilities in compiled binaries but permits source code analysis for defensive purposes.

To reduce disruption when safeguards do trigger, Anthropic introduced Automatic Fallbacks, a beta feature that routes blocked requests to a less powerful model instead of returning an error. Like its predecessor, Opus 5 is free from the 30-day data retention policy that applies to Fable and Mythos.

Enterprise and developer focus

Anthropic is marketing Opus 5 to enterprise customers for knowledge work, biology, and everyday tasks, while reserving Fable 5 for ambitious long-horizon agent projects. The company emphasized that Opus 5 is "much stronger at verifying its work and iterating carefully," citing examples where the model built a computer vision pipeline from an incomplete prompt.

The rollout comes amid heightened AI regulation. Following Fable 5's temporary takedown and negotiations with the Trump administration, OpenAI's GPT-5.6 underwent a limited preview restricted to government-approved entities. Anthropic says it continues to work with government partners on independent testing of Opus 5.