Claude Fable 5: Anthropic's Mythos-Class Model Goes Public

Claude Fable 5 launched on June 9, 2026 as Anthropic's first public release of Mythos-class capabilities. Until now, the Mythos tier was restricted to Project Glasswing partners for defensive cybersecurity. Fable 5 makes those capabilities available to all developers through the standard Claude API, with safety mechanisms that Anthropic describes as sufficient for general deployment. It currently ranks #1 out of 320 coding models with a composite score of 97/100.

The core distinction between Fable 5 and its predecessor Mythos is the safety layer: Fable 5 wraps Mythos-class intelligence in a deployment-safe architecture that keeps safeguards triggering in fewer than 5% of sessions, while maintaining zero compliance on adversarial jailbreak attempts. Anthropic's framing is explicit - this is Mythos for everyone, with the guardrails to make that responsible.

Fable 5's benchmark profile is the strongest Anthropic has ever published. The system card reports five headline results, each representing a new high-water mark for the Claude model family:

The standout result is SWE-bench Verified at 95% - a 14-point jump over Opus 4.8. SWE-bench tasks require models to navigate real GitHub repositories, understand thousands of files, locate bugs, and produce working patches. A 95% pass rate means Fable 5 solves nearly every real-world coding task the benchmark throws at it.

The HLE score of 59% deserves special attention. Humanity's Last Exam was specifically designed to resist AI progress - questions sourced from domain experts intended to stump frontier models. Fable 5's 19-point lead over Opus 4.8 suggests a qualitative leap in reasoning depth, not just incremental improvement. See our benchmarks explorer for per-benchmark rankings across all models.

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. This is premium pricing - roughly 67% more than Opus 4.8 on output. But the economics shift dramatically with prompt caching: a 90% discount on cached input tokens means repeated context (system prompts, codebases, documentation) costs just $1/M tokens after the first call.

For teams doing agentic coding where the same codebase context is passed repeatedly, the effective cost drops significantly. The 90% prompt cache discount is the most aggressive Anthropic has offered and makes Fable 5 economically viable for high-volume use cases that would be prohibitive at list price. Compare pricing against all models on our pricing calculator or see the Anthropic API pricing breakdown.

The central engineering challenge of Fable 5 was taking Mythos-class capabilities - including the cybersecurity skills that kept the original Mythos restricted - and making them safe for unrestricted API access. The system card details a multi-layered safety architecture:

The practical takeaway for developers: Fable 5 should behave like a capable, cooperative coding assistant in virtually all legitimate use cases. The safety layer is designed to be invisible during normal operation and only activates at the boundaries of Anthropic's usage policy. This contrasts with earlier Claude models where safety interventions were a more frequent friction point for developers.

Anthropic's system card includes real-world deployment results from enterprise partners. These are self-reported by Anthropic but provide concrete performance claims beyond synthetic benchmarks:

These case studies align with the benchmark data - a model that solves 95% of SWE-bench tasks should be capable of large-scale production engineering work. The Stripe result in particular suggests Fable 5 can handle the kind of marathon agentic sessions (thousands of sequential edits across thousands of files) that previously caused models to lose coherence.

Fable 5 is not a universal replacement for Opus 4.8. Its premium pricing means you should be deliberate about when to deploy it. Here is the decision framework:

The optimal strategy for most teams will be a tiered approach: use Fable 5 for the hard problems where its superior reasoning justifies the cost, and route simpler tasks to Opus 4.8 or lighter models. Anthropic's own routing architecture (the 95%/5% fallback split) models this approach. Compare all Anthropic models on our Claude models page.

Here are the current top 10 coding models, with Fable 5 highlighted. Rankings update hourly based on live benchmark data across SWE-bench, GPQA, HumanEval, MMLU, and 15+ evaluations.

Score bracket: 10 models score within 5 points of Fable 5, including Claude Opus 5 (Fast) (95), Claude Opus 5 (95), Claude Opus 4.8 (Fast) (95), and 7 others. Use our compare tool for head-to-head breakdowns.

Anthropic now has 21 models in our coding leaderboard. Here is the full Claude lineup ranked by score:

The biggest relative gains are on the hardest benchmarks: SWE-bench Pro jumps from 53.4% to 80.3% (a 50% relative improvement), and HLE from 40.0% to 59.0% (a 48% jump). These are the evaluations designed to separate frontier models from their predecessors, and Fable 5 clears the bar decisively. For a detailed comparison, use our compare tool.