Anthropic released Fable 5 to everyone on June 9, while keeping Mythos 5 restricted to a small group of trusted partners. The two models share the same underlying architecture — the difference is the safety layer. Fable 5 ships with guardrails; Mythos 5 has many of them removed, because the unrestricted model is considered too dangerous for open release.
How the Guardrails Work
Fable 5 adds runtime classifiers — independent AI systems that inspect every request. When a request touches cybersecurity, biochemistry, or asks the model to help build stronger AI, the workload is rerouted to the older Opus 4.8. The rerouting is relatively rare: on average, fewer than half of sessions are affected, and users mostly see a notification when it happens.
Mythos 5 is provided to Project Glasswing partners, who primarily work on defending critical software. Giving them the full version lets them find vulnerabilities and harden systems faster.
Capability Highlights
The model handles long-running, messy tasks that previously stalled within days:
- Stripe used it to process a 50-million-line Ruby codebase — work the team estimated would take two months — in a single day.
- It completed Pokémon purely from raw game pixels, with no map or state overlays, and planned automated factory lines in Factorio.
- In Slay the Spire, persistent file memory gave it a 3x higher clear rate than Opus 4.8.
- It built a solar system simulation from first-principles physics and predicted eclipses.
- Mythos 5 designed protein complexes autonomously; internal testing showed drug design workflows accelerated roughly 10x, with candidate molecules now under validation.
- It conducted genomics research: assembling data from millions of cells across 138 animal species, then training a small model to identify cell functions that outperformed a published Science model at 1% of its size.
- Capability cliffs. When classifiers trigger, performance drops back to Opus-level. Legitimate security research or bioinformatics queries may be misjudged.
- Restrictions on frontier LLM development. The system card states that if you ask Fable 5 to help build training pipelines, distributed training infrastructure, or ML accelerator designs, it will deliberately degrade results — not refusing outright, but quietly intervening via prompt modification and steering vectors. This is technical enforcement of anti-proliferation policy.
- Data retention. Fable 5 prompts and outputs are retained for 30 days to run the classifiers. Not used for training, but a privacy consideration.
- Selective access. Mythos 5 is first available to US-affiliated partners for infrastructure defense — no longer purely commercial behavior.
Pricing and Availability
Pricing is $10 per million input tokens and $50 per million output tokens — less than half the cost of Mythos Preview. It is available on claude.ai paid plans, the API, AWS Bedrock, Azure, and GitHub Copilot, with free trial periods on some plans.
The Caveats
Takeaways
For developers, long-horizon agents, complex coding, and vision-related tasks can now run far longer under model control. But the gate can close at any time, especially if you study the models themselves. The industry direction is shifting: instead of training a maximally safe model, Anthropic trains the strongest model and manages risk with runtime classifiers and tiered access.
Short term, trigger rates are low and coverage seems reasonable — but determined actors keep advancing jailbreak techniques. Long term, dual-track access may widen the gap between ordinary developers on the gated version and institutions with full access.
If you work on agent development, long-context applications, or scientific computing, Fable 5 is worth trying — especially for tasks requiring days of continuous operation. But don't put critical paths entirely on it: the provider's gate can change at any moment, so keeping your own toolchain as a fallback is always prudent.
Anthropic's new blade is fast — and sheathed. How and when it's drawn ultimately depends on the person holding it.