2 weeks ago
5542
AI giant Anthropic has on Monday released Claude Fable 5, a general-access version of its Mythos-class AI, which the company claims outperforms every model it has previously made publicly available. In addition, a restricted variant of the Mythos AI called Claude Mythos 5 will ship to US government cyber defenders through the existing Project Glasswing program.
Claude Fable 5 is built on the same AI architecture as Mythos but includes safety classifiers that route certain sensitive queries to Claude Opus 4.8 instead. Anthropic has acknowledged that the classifiers are tuned conservatively and will flag some harmless requests, however, the company has said this would happen in fewer than 5% of sessions on average, according to Anthropic’s website announcement.
The restricted Mythos 5 will instead direct these classifiers to specific domains. It comes as an upgrade to the Claude Mythos Preview that Glasswing participants have been using and will eventually reach a bigger set of already vetted customers, Anthropic said.
Claude Fable 5 performance indices
Anthropic positioned Fable 5 as the topmost performer on most capability benchmarks it tested, stating the model was further advantaged at longer and more complex tasks. The company highlighted several early-access results, which included multiple different tests at different firms.
Stripe reported that Fable 5 handled a codebase-wide migration across 50 million lines of Ruby in a single day, work the company estimated would have taken a full team more than two months manually. Cognition’s FrontierCode evaluated which models produced code that met production-quality standards the best, and concluded Fable 5 scored highest among frontier models, even at medium effort.
In finance, Hebbia’s tested benchmarks for senior-level analytical reasoning ranked Fable 5 first, with Anthropic citing performance gains in document reasoning and chart interpretation. Trading firm IMC said that in its own evaluations, the model performed exceedingly well at analyzing expected value, fact-checking, and conceptual reasoning.
Anthropic also claimed Fable 5 completed Pokémon FireRed using only raw game screenshots and a minimal harness on Vision, a task that earlier Claude models needed additional tooling to attempt.
On cybersecurity, Anthropic described Mythos 5 as having “the strongest cybersecurity capabilities of any model in the world,” though that claim is self-assessed and has not been independently verified through public benchmarks.
Pricing and availability
Both public and restricted models went live on Tuesday, June 9, 2026, and were priced at $10 per million input tokens and $50 per million output tokens, respectively. This rate is less than half of what Anthropic charged for the Claude Mythos Preview.
Fable 5 is accessible through the Claude API, Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry, according to Anthropic’s developer documentation. Mythos 5 remains limited to only approved Project Glasswing participants.
Both models support context windows of one million tokens, and can produce up to 128,000 output tokens per request, according to the API documentation.
Anthropic’s safety tradeoffs
Releasing a model at this capability level carries risks that Anthropic has openly discussed and acknowledged. Without the new model’s safety classifiers, Fable 5’s cybersecurity knowledge could be misused, the company said. The classifier system acts as a gatekeeper, giving users responses from Claude’s Opus 4.8 instead whenever it is triggered.
The API returns a specific `stop_reason: “refusal”` signal so developers can detect and handle these cases, according to the developer documentation.
Developers can configure automatic fallback to another Claude model when Fable 5 refuses a request, either server-side through a `fallbacks` parameter or on the client-side via SDK middleware, the documentation also states. Anthropic does not bill for requests refused before output generation and offers fallback credits to offset the cache cost of retrying prompts on a different model.
The company said it is working to reduce safety trigger false positives as more capable models arrive in the coming months.
The smartest crypto minds already read our newsletter. Want in? Join them.
English (US)