Claude Fable 5 vs Mythos 5: What the Safety Fallback Means

Claude Fable 5 and Claude Mythos 5 are the same base model with identical weights, the same 1M token context window, and the same $10 input / $50 output per million token pricing. The only real difference is safety: Fable 5 runs inference-time safety classifiers that fall back to Opus 4.8 on under 5% of requests, while Mythos 5 has no classifiers and sits behind a US government-linked approval program. As of June 28, 2026, neither model is available to the general public.

---

---

The short version: same brain, different guardrails

A "classifier" here is a small filter that reads a request before the main model answers. If the request looks dangerous, the classifier blocks it. That filter is the entire difference between these two models.

Fable 5 ships with those filters turned on. Mythos 5 ships with them turned off. Same weights underneath. So when you see a benchmark gap between Fable 5 and Mythos 5, you are not looking at a smarter model. You are looking at what the filters refuse to let Fable 5 do.

Why this matters to you: if you ever get access, you do not need to "pick the better model." You need to know what Fable 5 blocks and whether your workflow trips those blocks.

Fable 5 vs Mythos 5: full comparison

How the Fable 5 safety fallback actually works

This is the part most write-ups skip. When a Fable 5 classifier fires, the API does not throw an error. It returns a normal HTTP 200 response (the "everything is fine" status code) with stop_reason: "refusal". The response body tells you which classifier triggered. You are not billed for the refused request.

There is also a beta fallbacks parameter. Turn it on, and when a classifier fires, Anthropic's servers quietly retry your request on Opus 4.8 instead of failing. If you paid for prompt caching (storing part of your prompt so repeat calls are cheaper), a "fallback credit" refunds that cost so you are not charged twice.

In plain terms: Fable 5 rarely says no, and when it does, it either hands the job to an older model or gives you your money back.

The three things that trigger a Fable 5 refusal

Anthropic built the classifiers around three areas. If your work touches these, expect refusals at runtime, so design around them early:

1. Offensive cyber and exploitation. Writing or finding software attacks. This is the big one. On ExploitBench, Mythos 5 scores 78% versus Opus 4.8 at 40%. That gap is the whole reason the classifier exists.
2. Gene therapy design and dual-use biotech. Anything that could double as a bioweapon recipe.
3. Model distillation. Trying to copy the model by harvesting its outputs to train a competitor.

If your app does normal product work (auth flows, payments, dashboards), you will almost never see these. If you build security tooling, plan for the fallback.

Why Mythos 5 is locked and who gets it

Mythos 5 has no safety filters, so Anthropic does not sell it openly. Access runs through Project Glasswing, a US government-linked program for vetted cyber defenders and critical infrastructure operators (think power grids, water systems, hospitals). Standard API customers cannot apply through normal channels.

The 78% ExploitBench number is the public justification. Anthropic released it alongside the launch to explain why a no-filter model is not offered to everyone. A model that can find real software attacks 78% of the time is a defensive tool for the right hands and a weapon in the wrong ones.

The suspension: why you cannot use either one

Here is the timeline as of June 28, 2026:

• Both models launched in June 2026.
• Three days later, a reported jailbreak surfaced. A "jailbreak" is a trick prompt that gets a model to ignore its safety rules.
• On June 12, 2026, a US export-control directive suspended both models worldwide. Anthropic called the jailbreak "narrow and non-universal" and noted (reported) that the same trick also works on GPT-5.5, but complied anyway.
• Fable 5 has stayed dark for 16 days, with no announced return date.
• On June 27, 2026, Mythos 5 was partially restored under "Mythos Annex A" for US critical infrastructure operators only.

So the most discussed Anthropic comparison of the year is one almost nobody can run.

The performance that justifies the price

When live, the base capability is strong. SWE-Bench Pro (a test of fixing real software bugs) puts Fable 5 at 80.3% versus GPT-5.5 at 58.6%. On the FrontierCode Diamond Split (the hardest coding problems), Fable 5 hits 29.3% versus GPT-5.5 at 5.7%. Anthropic's Stripe case study reports a 50-million-line Ruby migration finished in one day, against a manual estimate of two-plus months.

These numbers describe the shared base model, so they apply to Mythos 5 too. The classifiers do not make Fable 5 worse at coding. They only block the three risk areas above.

Constraints both models share

Whatever tier you land on, three limits apply and matter for compliance reviews:

• 30-day data retention is mandatory. Zero data retention is not offered. Your prompts are stored for 30 days.
• Adaptive thinking is always on. You cannot pass thinking: disabled.
• Raw chain-of-thought is never returned. You see the answer, not the model's full internal reasoning.

If your industry requires zero retention, neither model fits today.

Where Build This Now fits

You do not need a frontier-tier model to ship a real product. The $29 Code Kit is a build system for Claude Code: a production SaaS skeleton with auth, Stripe payments, and PostgreSQL with row-level security on every table, plus agents, skills, and CLAUDE.md workflows wired together so Claude Code ships full apps instead of snippets. It runs on whatever Claude model your subscription gives you, so you build today and swap in stronger models later. Pair it with Claude Code subagents and MCP servers and one person can do the work of a small team.

FAQ

Are Claude Fable 5 and Mythos 5 the same model?

They share identical base weights. The only difference is that Fable 5 adds inference-time safety classifiers that fall back to Opus 4.8 on a small share of requests, while Mythos 5 has no classifiers and is restricted to vetted partners.

Why can't I access Claude Mythos 5?

Mythos 5 is gated behind Project Glasswing, a US government-linked approval program for vetted cyber defenders and critical infrastructure operators. Standard API customers cannot apply through normal channels.

Is Claude Fable 5 available yet?

As of June 28, 2026, Fable 5 is suspended worldwide due to a US export-control directive issued June 12, 2026. No return date has been announced. Mythos 5 was partially restored June 27 for US critical infrastructure operators only.

What happens when Claude Fable 5 refuses a request?

The API returns stop_reason: "refusal" as an HTTP 200 response, not an error, names which classifier fired, and does not bill you for the refused request. With the beta fallbacks parameter, Anthropic retries the request on Opus 4.8 server-side, and a fallback credit refunds any prompt-cache cost so you are not double-billed.