Claude Fable 5 & Mythos 5: Anthropic's Most Capable Models Yet

On June 9, 2026, Anthropic introduced its most capable models to date: Claude Fable 5 and Claude Mythos 5. They share one underlying model but ship as a deliberate pair — Fable 5 is the general-availability version with full safeguards, while Mythos 5 is the same model with certain protections removed for a small set of vetted users. For anyone building AI agents, the headline is not just raw benchmark wins; it is how far these models push long-horizon autonomy — staying coherent and useful across tasks that run for hours and span millions of tokens.

Key takeaways

• Two models, one brain: Fable 5 (safeguarded, general availability) and Mythos 5 (same model, fewer restrictions, trusted partners only).
• Pricing: $10 per million input tokens and $50 per million output tokens — less than half the cost of the earlier Mythos Preview.
• API model ID: claude-fable-5, available immediately on the Claude API and consumption-based Enterprise plans.
• Frontier results in software engineering, knowledge work, vision, and long-context reasoning — with the lead widening as tasks get longer and harder.
• Safety by routing, not refusal: risky cyber requests fall back to Claude Opus 4.8; safeguards trigger in under 5% of sessions.

What are Claude Fable 5 and Mythos 5?

Claude Fable 5 is the model most teams will actually use: it is available through the standard API and enterprise plans, and it ships with a layered safety system designed to keep frontier capability broadly accessible without handing dangerous uplift to bad actors. Claude Mythos 5 is the identical model with several of those guardrails lifted, made available only to restricted users such as Project Glasswing cybersecurity partners, infrastructure providers, and a select group of biomedical researchers operating under specific access programs. In plain terms: same intelligence, two different risk postures.

This dual-model approach is the interesting design decision. Rather than ship one watered-down public model and keep the real thing internal, Anthropic exposes the same capability ceiling to everyone via Fable 5 and uses classifiers to gate the genuinely sensitive requests — reserving the unrestricted Mythos 5 for partners who have been vetted and who accept extra conditions like a 30-day data-retention policy with logged human access.

How the safeguards actually work

The most builder-relevant detail is that Fable 5 treats safety as a routing problem. Incoming requests pass through three classifier categories — cybersecurity, biology/chemistry, and distillation prevention. A request flagged as offensive-cyber does not get a flat refusal; it quietly falls back to Claude Opus 4.8, so legitimate security work keeps flowing while exploitation attempts lose the frontier edge. Biology and chemistry requests that could provide uplift are blocked outright, and large-scale attempts to extract or distill the model's capabilities are stopped. Anthropic reports these safeguards engage in fewer than 5% of sessions, meaning over 95% of Fable 5 traffic runs at full capability with no fallback at all.

flowchart TD
  A["Agent or user calls claude-fable-5"] --> B{"Safety classifiers screen the request"}
  B -->|Offensive-cyber detected| C["Fallback to Claude Opus 4.8 (no hard refusal)"]
  B -->|Biology / chemistry uplift risk| D["Response blocked"]
  B -->|Bulk distillation attempt| E["Request blocked"]
  B -->|Clean: 95%+ of sessions| F["Full Fable 5 capability runs"]
  F --> G["Long-horizon work: code, analysis, vision, science"]
  G --> H["Result returned to the user or agent"]

Where it pulls ahead: long-horizon work

Benchmarks tell a consistent story — the advantage grows with task complexity and duration. A few concrete data points Anthropic shared:

• Software engineering: Stripe used Fable 5 to migrate a 50-million-line Ruby codebase in a single day — work that would have taken a full team over two months by hand. It also posted the top score among frontier models on Cognition's FrontierCode evaluation.
• Knowledge work: highest score on Hebbia's Finance Benchmark; one trading firm said it "aced" their trading-analysis evaluations almost across the board.
• Vision: completed Pokémon FireRed from raw screenshots with minimal scaffolding, where earlier Claude models needed elaborate harnesses.
• Long context: in Slay the Spire tests, persistent file memory boosted performance roughly 3x more than Opus 4.8, and it reached the final act 3x more often.
• Science (Mythos 5): ~10x faster protein-design iterations (9 of 14 targets produced strong candidates), novel molecular-biology hypotheses preferred ~80% of the time by scientists in blinded comparisons, and an autonomous week-long genomics run that trained a model 100x smaller than a published one while beating it.

Pricing and availability for builders

Both models are priced at $10 / million input tokens and $50 / million output tokens. Fable 5 is live now on the Claude API and consumption-based Enterprise plans. On subscriptions, it is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost from June 9–22, 2026; on June 23 it is removed from those subscription plans, with standard-plan access expected to return as capacity allows (it stays available via the API and usage-based plans). Mythos 5 remains gated to trusted partners, expanding gradually through Anthropic's trusted-access and biology-access programs.

Fable 5 vs. Mythos 5 at a glance

Common pitfalls when adopting a frontier model like this

• Treating it like a chatbot. The gains here are in multi-hour, multi-step autonomy. If your harness resets context every few turns, you leave most of the value on the table — give the agent durable memory.
• Ignoring the fallback. Some security-adjacent prompts will be served by Opus 4.8 instead of Fable 5. Detect the capability shift rather than assuming every call is the frontier model.
• Output-token sprawl. At $50/M output, long autonomous runs can get expensive fast. Cap reasoning where you can and cache aggressively.
• Skipping evals. "State of the art" on public benchmarks does not guarantee wins on your workload. Keep a small, honest eval set and gate releases on it.

What it means for AI agents

The throughline of this release is that frontier models are getting dramatically better at the thing agents need most: holding a goal steady across long, messy, tool-heavy work. Cheaper pricing plus stronger long-horizon reasoning means the kinds of multi-agent workflows that were flaky a year ago — research, migrations, end-to-end operations — are becoming dependable. That is the same shift that makes production voice and chat agents viable in the real world.

Frequently asked questions

What is the difference between Claude Fable 5 and Mythos 5?

They are the same underlying model. Fable 5 is the public, fully safeguarded version; Mythos 5 is restricted to vetted partners with certain protections removed for sensitive research and defense work.

How much does Claude Fable 5 cost?

$10 per million input tokens and $50 per million output tokens — less than half the price of the earlier Claude Mythos Preview.

What is the API model ID?

It is claude-fable-5, available on the Claude API and consumption-based Enterprise plans immediately.

Does Fable 5 refuse risky requests?

Not usually. Offensive-cyber requests fall back to Claude Opus 4.8 rather than getting a hard refusal, and the safeguards trigger in under 5% of sessions overall.

Bringing frontier agentic AI to your phone lines

CallSphere applies these same long-horizon, multi-agent patterns to voice and chat — AI agents that answer every call and message, use tools mid-conversation, and book work 24/7. See it live at callsphere.ai.

Source & attribution: This is an independent, original explainer inspired by Anthropic's announcement. Claude, Claude Fable 5, Claude Mythos 5, and Claude Opus are products and trademarks of Anthropic. CallSphere is not affiliated with or endorsed by Anthropic.