Anthropic 推出 Claude Fable 5 與 Claude Mythos 5：相同底層模型、不同安全防護，全新 Mythos 級別

Anthropic released two models on June 9, 2026: Claude Fable 5 and Claude Mythos 5. Both belong to a tier called “Mythos-class.” This tier sits above the Opus class in capability. Fable 5 is the version claimed to be made safe for general use. Mythos 5 is the same model with some safeguards lifted, kept in limited release. Claude Fable 5 and Mythos 5 Mythos-class models are a tier of Claude models. They sit above the Opus class in capability. The first was Claude Mythos Preview, released in April through Project Glasswing. Fable 5 and Mythos 5 share the same underlying model. The difference is the safeguards. Fable 5 ships with safety classifiers for general use. Mythos 5 has some classifiers removed and stays in limited release. The names reflect this split. “Fable” comes from the Latin fabula, “that which is told.” This is akin to the Greek mythos. The safeguards distinguish the two models, so they carry different names. Anthropic team calls Fable 5 its most capable widely released model. It targets demanding reasoning and long-horizon agentic work. Anthropic states Fable 5’s capabilities exceed any model it has made generally available. Both models support a 1M token context window by default. They allow up to 128k output tokens per request. Pricing is $10 per million input tokens and $50 per million output tokens. That is less than half the price of Claude Mythos Preview. The Capability Case Anthropic reports Fable 5 is state-of-the-art on nearly all tested capability benchmarks. It shows strong results across software engineering, knowledge work, vision, and scientific research. The longer and more complex the task, the larger its lead over Anthropic’s other models. On software engineering, Stripe tested Fable 5 during early access. The model performed a codebase-wide migration in a 50-million-line Ruby codebase. According to Stripe: this took one day. By hand, a team would have needed over two months. Fable 5 is also more token-efficient than past Claude models. On Cognition’s FrontierCode evaluation, Fable 5 scores highest among frontier models. This holds even at medium effort. The eval tests difficult coding tasks under production-codebase standards. On knowledge work, Anthropic cites Hebbia’s Finance Benchmark for senior-level reasoning. Fable 5 posts the highest score of any model there. Gains come in document-based reasoning, chart and table interpretation, and problem solving. On vision, Anthropic calls Fable 5 the new state-of-the-art. It can extract precise numbers from detailed scientific figures. It can rebuild a web app’s source code from screenshots alone. It also needs less scaffolding than prior models. Fable 5 beat Pokémon FireRed with a minimal, vision-only harness. On memory and long-context, Fable 5 stays focused across millions of tokens. It improves its outputs using its own notes. In the game Slay the Spire, persistent file-based memory helped it three times more than Opus 4.8. Mythos 5 carries the science claims. Internal protein design experts accelerated parts of drug design by around ten times. Anthropic also says Mythos 5 is its first model to consistently produce novel scientific hypotheses. Scientists preferred its molecular biology hypotheses around 80% of the time in blinded comparisons. Mythos 5 also ran novel genomics research over a week of largely autonomous work. It trained a custom model on single-cell data spanning 138 animal species. Anthropic says that model outperformed a recent model published in Science, despite being 100 times smaller. How the Safeguards Work Releasing a model this capable carries risk. Without safeguards, Fable 5’s cybersecurity capabilities could be misused to cause serious damage. Anthropic therefore launched Fable 5 with a new set of classifiers. Classifiers are separate AI systems. They detect potential misuse, including jailbreak attempts. They prevent the main model from responding to flagged requests. When Fable 5’s classifiers flag a request, the response is handled by Claude Opus 4.8 instead. The covered areas are cybersecurity, biology and chemistry, and distillation. Users are informed whenever a fallback occurs. For biology and chemistry, Fable 5 falls back to Opus 4.8 on most requests for now. Anthropic cites concern that the same dual-use queries could give uplift to malicious actors. It plans a trusted access program for biology, giving approved researchers Fable 5 without those safeguards. Anthropic tuned these safeguards conservatively. They will sometimes catch harmless requests. On average, they trigger in less than 5% of sessions. Anthropic says more than 95% of Fable sessions involve no fallback at all. For those sessions, Fable 5’s performance effectively matches Mythos 5. Anthropic red-teamed the classifiers extensively. An external bug bounty produced no universal jailbreaks in over 1,000 hours. A universal jailbreak lets a user interact with the model as if its safeguards were absent. Anthropic notes the UK AISI made progress toward one in a brief testing window. Mythos 5 is the same model with cyber safeguards lifted. Anthropic describes it as having the strongest cybersecurity capabilities of any current model. It is deployed through Project Glasswing in collaboration with the US government. Use Cases These capabilities map to several concrete workflows for technical teams: Large-scale code migration: Long-horizon coding suits big refactors and cross-repo migrations. The Stripe example shows this at a 50-million-line scale. Agentic coding pipelines: Fewer turns and token efficiency help multi-step agent runs. GitHub reported autonomy and reliability on complex, long-horizon coding tasks. Finance and analytics work: Strong document and chart reasoning suits senior-level financial analysis. Hebbia and IMC cited gains on reasoning and trading-analysis tasks. Vision-to-code tasks: Rebuilding source from screenshots suits front-end reconstruction and figure extraction. The vision-only harness reduces tooling overhead. Long-running research agents: Persistent memory across millions of tokens suits multi-day research loops. Mythos 5 ran novel genomics work over a week of largely autonomous work. Comparison Table: Fable 5 vs. Mythos 5 vs. Opus 4.8 AttributeClaude Fable 5Claude Mythos 5Claude Opus 4.8Model tierMythos-classMythos-classOpus classUnderlying modelSame as Mythos 5Same as Fable 5Opus 4.8AvailabilityGenerally availableLimited (Project Glasswing)Generally availableSafety classifiersActive (cyber, bio/chem, distillation)Cyber safeguards liftedOpus-level safeguardsFallback targetFalls back to Opus 4.8Not applicableNot applicableAPI model IDclaude-fable-5claude-mythos-5(existing Opus ID)Context window1M tokens default1M tokens defaultPer Opus specsMax output128k tokens/request128k tokens/requestPer Opus specsInput price (per 1M)$10$10(per Opus pricing)Output price (per 1M)$50$50(per Opus pricing)Thinking modeAdaptive only, always onAdaptive only, always onConfigurableData retention30-day (Covered Model)30-day (Covered Model)Standard options Note: Specific Opus 4.8 specs and pricing are not detailed in the Fable 5 launch sources. The table marks those cells accordingly. Key Takeaways Fable 5 and Mythos 5 share one underlying model; safeguards are the only difference. Anthropic reports Fable 5 is state-of-the-art on nearly all tested capability benchmarks. Fable 5 classifiers fall back to Opus 4.8 and trigger in under 5% of sessions. Both models offer a 1M token context window at $10 input and $50 output per million tokens. Mythos 5 stays limited to Project Glasswing; Fable 5 is generally available across major platforms. Sentiments of the Community #mtp-mythos-sentiment *{margin:0;padding:0;box-sizing:border-box!important} #mtp-mythos-sentiment hr,#mtp-mythos-sentiment p:empty,#mtp-mythos-sentiment del,#mtp-mythos-sentiment s{display:none!important} #mtp-mythos-sentiment br{display:none!important} #mtp-mythos-sentiment{ --bg:#0F0F0F;--bg2:#161616;--line:#24242