Overview
Claude Fable 5 is Anthropic's flagship new model released on June 9, 2026, and is the company's **first publicly released Mythos-level model**, marking Claude's official generational leap to the 5 series—distinct from previous minor version iterations of 4.x.
Fable 5 shares the **exact same underlying model** as the Mythos 5 released on the same day, with the only difference being that Fable 5 adds three layers of safety guardrails (cybersecurity/biosecurity/distillation protection). Requests that hit these guardrails are automatically forwarded to Opus 4.8 for processing (in reality, fewer than 5% of conversations trigger this). Anthropic officially states that it "achieves SOTA on almost all benchmarks, with greater advantages on longer and more complex tasks," and Karpathy calls it "a major leap forward."
Key Features
- Codebase-scale Ultra-large Engineering: Stripe test: Migrating a 50-million-line Ruby codebase compressed from 2 months to 1 day, with token efficiency surpassing all previous Claude models
- FrontierCode / CursorBench SOTA: Both Cognition FrontierCode and the Cursor team rate it as the highest score among frontier models (achievable with moderate inference intensity)
- Finance and Knowledge Work SOTA: Highest scores across all Hebbia financial benchmarks, with near-universal excellence in IMC trading analysis (factual queries/conceptual reasoning/root cause analysis/expected value)
- Leap in Pure Vision Capabilities: Rebuilds web application source code from screenshots alone; uses a pure vision harness to beat Pokémon: FireRed (previous generations couldn't even with tools)
- Maintains Focus on Multi-million Token Long Tasks: In Slay the Spire tests with persistent file memory, reaches the final chapter 3 times more frequently than Opus 4.8
- Triple Safety Guardrails + Opus 4.8 Fallback: Three types of sensitive requests (cybersecurity/biosecurity/distillation) are automatically routed to Opus 4.8; over 95% of conversations do not trigger this
Use Cases
- Ultra-large codebase migration and refactoring (million-line scale)
- Financial analysis, complex document reasoning (charts/tables/expected value analysis)
- Pure vision-driven UI reconstruction and interface understanding
- Multi-million token long tasks (data analysis, long-form writing, Agent long-horizon tasks)
- Enterprise production scenarios requiring top-tier reasoning + engineering-level stability
Pros
- Generational improvement in general capabilities, SOTA on almost all benchmarks
- Greatest advantage in engineering tasks (codebase-scale)
- Breakthrough in pure vision capabilities, achieving screenshot reconstruction and visual game completion for the first time
- Pricing less than half that of Mythos Preview ($10/$50 per million tokens)
- Alignment performance on par with Opus 4.8 (low level of misalignment behavior)
Pricing
API pricing: $10/million tokens for input, $50/million tokens for output, roughly half the price of Claude Mythos Preview. Subscription availability is phased: from 6/9 to 6/22, Pro/Max/Team/per-seat Enterprise plans include Fable 5 for free; starting 6/23, it is removed from subscriptions and requires usage credits (extended if capacity allows); once capacity is sufficient, it will be restored as a standard subscription component. Claude API and pay-as-you-go Enterprise are fully available immediately, with the API call name `claude-fable-5`.
Summary
Fable 5 suits two types of users: engineering teams handling ultra-large codebase refactoring/full codebase migration tasks (where its advantage is most pronounced); and finance/knowledge workers processing long documents and complex reasoning. For everyday chatting, email writing, or small research tasks, the higher-priced Fable 5 may not be cost-effective—Sonnet 4.8 is more economical. The Mythos-level guardrail mechanism is imperceptible for most daily tasks, but sensitive areas like biosecurity and cybersecurity will automatically fall back to Opus 4.8—this is the core implementation of its "safe and controllable Mythos-level" positioning.