Anthropic Releases Claude Fable 5 for Broad Use

Anthropic announced the general release of Claude Fable 5 on June 9, 2026, and the limited-availability roll out of Claude Mythos 5 via Project Glasswing, per Anthropic's blog and platform documentation. Anthropic's product page lists the API model as claude-fable-5, supports up to 128k output tokens per request, and shows pricing at $10 per million input tokens and $50 per million output tokens. GitHub's changelog and Amazon's product announcement both report that Claude Fable 5 is available in GitHub Copilot and on AWS Bedrock and other cloud integrations.

Anthropic's platform documentation and system card describe Claude Fable 5 and Claude Mythos 5 as configurations of the same underlying Mythos-class model family. The published material highlights longer-horizon, agentic capabilities such as extended asynchronous execution, advanced vision understanding for diagrams and nested tables, and self-verification behaviors. The platform docs document API behaviour for refusals, fallback handling via a fallbacks parameter, and SDK middleware for retrying on alternate models. GitHub's changelog notes internal benchmarking where Fable 5 completed equivalent autonomous coding workflows with fewer tool calls and lower token consumption than prior Opus-tier models.

Releasing a Mythos-class model broadly, while retaining a sealed higher-trust variant, mirrors a recent industry pattern of tiered frontier models combined with access controls. Anthropic's simultaneous emphasis on capability (Mythos-level performance across coding, vision, and long tasks) and operational safeguards (classifiers, refusal semantics, and explicit retention for safety tooling) aligns with public reporting by The Verge, TechCrunch, CNBC, and Wired that frame this launch as both capability-forward and safety-conscious. For enterprise consumers the trade-offs are now explicit: broader access to a higher-capability model comes with documented constraints on retention and classifiers that affect integration and compliance.

Teams evaluating claude-fable-5 should treat the model as a high-capability, higher-cost option. Reported pricing ($10/M input, $50/M output) and large-context support (up to 128k output tokens) will materially change token-cost calculations for long-running or agentic workflows. Anthropic reports its safety classifiers trigger in fewer than 5% of sessions on average; in those cases, the platform returns a Claude Opus 4.8 response rather than a hard refusal, meaning integration code must handle responses from a different model tier alongside any stop_reason: "refusal" rejections. GitHub's changelog explicitly notes that enabling Claude Fable 5 in Copilot requires turning on a policy that accepts up to 30 days of prompt/output retention; GitHub's post highlights that retained data is not used to train Anthropic models.

Observers should track:

• •how often the safety classifiers trigger refusals in real-world developer and enterprise workloads
• •empirical token and tool-call efficiencies claimed by GitHub benchmarks versus independent tests
• •uptake and behavior differences between the broadly available Fable 5 and the limited Mythos 5 instances available through Project Glasswing. Also monitor third-party benchmarking and security research (e.g., jailbreak and vulnerability results) because Mythos-class capability increases both utility and adversarial risk, a point emphasized in industry reporting