Anthropic has released Claude Fable 5 to the public, positioning the model as its most capable general-purpose system yet while adding safety controls intended to limit misuse.

The company said Fable 5 outperforms its earlier models across a wide range of benchmarks, with especially strong results in software engineering, analytical work, vision tasks, scientific research and long-running autonomous jobs. Anthropic described the model as a Mythos-class system that has been made safe for general use.

The launch also comes with limits. Anthropic said queries involving certain sensitive areas, including topics that could be used for cybersecurity abuse, will be routed to Claude Opus 4.8, its next-most-capable model. The company said those safeguards were tuned conservatively to reduce risk, though they may block some harmless prompts. According to Anthropic, the protections trigger in fewer than 5% of sessions on average.

Alongside Fable 5, Anthropic introduced Claude Mythos 5 for a smaller group of cyberdefenders and infrastructure providers. Mythos 5 uses the same underlying model but removes some safeguards in selected areas. It is initially being deployed through Project Glasswing, a collaboration with the U.S. government, as an upgrade to Claude Mythos Preview. Anthropic said Mythos 5 has the strongest cybersecurity capabilities of any model it has tested and plans to expand access through a broader trusted-access program.

Anthropic set pricing for both models at $10 per million input tokens and $50 per million output tokens, which it said is less than half the price of Claude Mythos Preview. The company framed the release as part of a broader effort to make advanced AI widely available while trying to do so safely.

In software engineering, Anthropic said early users reported that Fable 5 could handle large-scale tasks that previously took teams much longer. Stripe, one of the early testers cited by the company, said the model compressed months of work into days and completed a codebase-wide migration in a 50-million-line Ruby repository in a day. Anthropic also highlighted results from Cognition’s FrontierCode evaluation, where it said Fable 5 ranked highest among frontier models at medium effort.

The model also showed strong performance on knowledge work benchmarks, according to Anthropic. The company said it scored highest on Hebbia’s Finance Benchmark for senior-level reasoning and performed well on document analysis, chart interpretation and problem solving. IMC, another tester cited by Anthropic, said the model performed strongly in trading-analysis evaluations.

Vision was another area where Anthropic said the model made a notable leap. The company said Fable 5 can extract detailed figures from scientific images, rebuild web apps from screenshots and complete vision-only games such as Pokémon FireRed without extra navigation tools.

Anthropic also emphasized memory and long-context performance. It said the model can stay on task across millions of tokens and improve its work by referring to its own notes. In a game-testing example, persistent file-based memory materially improved performance in Slay the Spire.

Beyond software, Anthropic said Mythos 5 has already contributed to drug-discovery work, accelerating parts of protein design by about 10 times in internal tests. The company also said the model generated new hypotheses in molecular biology and carried out a week of autonomous genomics research.

Anthropic said its automated alignment tests found Mythos 5’s misaligned behavior was low and similar to Opus 4.8. The company said Fable 5 should show similar behavior because the two models share the same underlying system.