Anthropic Ships Fable 5 with Guardrails; Mythos 5 Stays Restricted

Forum topic · ✨步子哥 · 2026-06-10

Summary

Anthropic released Fable 5 to the public on June 9 while keeping its more capable sibling, Mythos 5, limited to trusted partners. Both models share the same underlying architecture; the key difference is safety controls. Fable 5 runs independent AI classifiers that route high-risk requests—cybersecurity, biochemistry, or frontier AI development assistance—to the older Opus 4.8, affecting on average fewer than half of sessions. Mythos 5 goes to Project Glasswing partners focused on critical software defense. Fable 5 handles long-running agent tasks, completed a 50-million-line Ruby codebase migration in a day versus an estimated two months, beat Pokémon and Factorio from raw visuals, designed protein complexes accelerating drug design roughly 10x, and produced a cell-function model outperforming a published Science model at 1% of its size. Pricing is $10 per million input tokens and $50 per million output tokens, less than half of Mythos Preview. Concerns include capability downgrades on classifier triggers, subtle restrictions on frontier LLM development work via prompt modification and steering vectors, 30-day prompt/output retention, and a widening gap between restricted public access and full partner access.

Anthropic released Fable 5 to everyone on June 9, while keeping Mythos 5 for a small group of trusted partners.

Same Skeleton, Different Locks

The two models share the same underlying architecture. The difference is the safety switch. Fable 5 ships with guardrails; on Mythos 5, many of those locks are removed. The reasoning: the model is powerful enough that unrestricted public access could enable high-risk misuse, so Anthropic did not want to release the full version to the open market.

Fable 5 adds new classifiers—independent AI systems that monitor every request. When a request touches cybersecurity, biochemistry, or asks the model to help build stronger AI, the task is routed to the older Opus 4.8. Reroutes are not frequent: on average fewer than half of sessions are affected, and users mostly see a notification when it happens.

Mythos 5 is reserved for Project Glasswing partners, who work mainly on defending critical software; full access lets them find vulnerabilities and secure systems faster.

Capabilities

Long-running tasks: Where previous models could not sustain multi-day work, Fable 5 can. The longer and messier the task, the bigger its advantage.
Coding: Stripe used it on a 50-million-line Ruby codebase—a job the team estimated at two months—completed in a day.
Vision: It beat Pokémon using only game footage, with no map or status readouts. It also planned and built automated factory lines in Factorio on its own.
Games: In Slay the Spire, persistent file-based memory gave it a 3x higher win rate than Opus 4.8.
Simulation: It drew its own solar-system simulation, deriving orbits from first principles and predicting eclipses.
Science: Mythos 5 designed protein complexes and accelerated drug-design pipelines roughly 10x in internal tests; for some targets it selected sites, ran tools, recovered from failures, and produced candidate molecules now under validation.
Genomics: It ran a genomics study assembling data on millions of cells across 138 animal species, then trained a small model to identify cell functions that outperformed a model published in Science while being only 1% of its size.

Tasks that once required scientists guiding every step can now run autonomously with only high-level instructions.

Pricing and Availability

Input: $10 per million tokens; output: $50 per million tokens—more than 50% cheaper than Mythos Preview.
Available via claude.ai paid plans, API, AWS Bedrock, Azure, and GitHub Copilot, with free trial access on some plans for a limited time.

The Caveats

Classifier triggers: Though rare, a trigger drops you back to Opus-level capability. For security researchers or bioinformaticians, legitimate questions may be misjudged.
Frontier LLM restrictions: The system card is explicit—requests for help building training pipelines, distributed training infrastructure, or ML accelerator designs are actively limited, not by refusal but quietly via prompt modification and steering vectors. This is proliferation control: Anthropic is enforcing its own counter-proliferation policy with technical means.
Data retention: Fable 5 prompts and outputs are retained for 30 days to run classifiers—not used for training, but an added privacy consideration.
Glasswing access: Prioritizing US-aligned partners for infrastructure defense is not purely commercial behavior.

What It Means

For developers, the upside is stronger tools for long-horizon agents, complex coding, and vision tasks. The downside is that the gate can close at any time—especially if you study models themselves.

The industry pattern has shifted: instead of training a safe model, Anthropic trains the strongest model and manages it with runtime classifiers and tiered access.

Short term: Low trigger rates and reasonable coverage—but determined actors keep advancing jailbreak techniques.
Long term: A two-track system (locked version for ordinary developers, full version for specific institutions) could tilt the competitive balance.

On the positive side, Anthropic did not lock everything down. The model and its safety mechanisms are both advancing—just toward runtime intervention plus access control, rather than alignment through training alone.

If you work on agent development, long-context applications, or scientific computing—especially tasks requiring days of uninterrupted autonomy—Fable 5 is worth trying. But don't put your entire critical path on it: the provider's gate can change overnight. Keeping your own toolchain as backup is always the safer play.

This new blade is sharp—and sheathed. How it gets drawn and used ultimately depends on the person holding it.

*Note: This is a forum post about an Anthropic release; specific benchmark claims reflect the original post and have not been independently verified.*