Explainer

Mythos-class, explained: why Anthropic shipped the same model twice

Fable 5 and Mythos 5 are the same underlying model, shipped under two different access policies. "Mythos-class" is Anthropic's name for a new capability tier it places above the Opus class; Fable 5 is the version of that model everyone can use, carrying safety classifiers, and Mythos 5 is the same model with those classifiers removed, restricted to approved organisations. The difference between them is not capability - it is who is allowed to run the unfiltered version.

Dated analysis. Checked 18 July 2026 against Anthropic's own documentation. This piece explains the tier and the classifier mechanics; the launch, the export-control suspension and the post-relaunch backlash are a separate story, covered in what launched, what broke, and what it costs.

What "Mythos-class" means

Anthropic describes Fable 5 as the first model in a new "Mythos-class" tier that sits above the existing Opus class in capability, built for demanding reasoning and long-horizon agentic work. (Anthropic, 9 June 2026; Claude Platform docs, checked 18 July 2026.) The name is a tier label, not a second model family: it tells you where the model sits relative to Opus 4.8, which now becomes the fallback rather than the flagship.

One model, two access policies

Anthropic states plainly that Mythos 5 is the same underlying model as Fable 5, minus Fable 5's safety classifiers, and that the docs put it as "Claude Mythos 5 shares Claude Fable 5's capabilities." The two share the same specifications: a 1,000,000-token context window, up to 128,000 tokens of output per request, adaptive thinking always on, and a price of $10 per million input tokens and $50 per million output tokens. The API model IDs are claude-fable-5 and claude-mythos-5. What separates them is a policy layer, not a training difference.

Fable 5 is generally available - on the Claude API, claude.ai, Claude Code, Amazon Bedrock, Google Cloud's Vertex AI and Microsoft Foundry. Mythos 5 is not: it is offered only in limited availability to organisations approved under Project Glasswing, Anthropic's restricted-access programme run in collaboration with the US government for approved cyberdefence, infrastructure and biology-research organisations. Access is brokered through an Anthropic, AWS or Google Cloud account team, not self-serve.

What the classifiers actually do

Fable 5's safety classifiers target three dual-use domains: cybersecurity, biology and chemistry, and suspected model-distillation attempts. When a classifier fires, Anthropic does not error the request - it answers it with Claude Opus 4.8 instead, and the API returns stop_reason: "refusal" as an HTTP 200 response, identifying which classifier triggered. In other words, a flagged query silently drops to a less capable model rather than failing outright. Anthropic states these safeguards trigger in fewer than 5% of Fable 5 sessions, meaning more than 95% of usage reaches the full model - a figure corroborated by secondary coverage. (Vellum.)

Mythos 5 has none of these classifiers, which is precisely why it is restricted. The point of the two-variant design is that the capability Anthropic is unwilling to expose to the open market in high-risk domains is exactly the capability approved defenders may need. Vendor-reported figures, relayed by a single secondary source, put numbers on that gap: on a hard biology-reasoning eval the unblocked model scored 46.1% against Opus 4.8's 40.0%, and on a cybersecurity capture eval it scored 78.0% against 40.0%.

Vendor-reported "safeguards removed" figures, relayed by a single secondary source (Vellum); illustrative of what the classifiers gate, not independently verified.
EvaluationMythos 5 (unblocked)Opus 4.8
BioMysteryBench hard biology reasoning46.140.0
ExploitBench cybersecurity capture78.040.0

Treat those two rows with the caution they deserve: they are vendor numbers, passed through one aggregator, on evals whose harnesses are not described here. They are useful only as an illustration of why the unfiltered variant is access-controlled, not as a benchmark you should trust to two decimal places.

What this means in practice

For almost everyone, the practical model is Fable 5, and the practical thing to know is the fallback: in under 5% of sessions - those touching cybersecurity, bio/chem, or that look like distillation - you are answered by Opus 4.8, flagged as a refusal, not by the model you paid for. If your work lives in those domains, that fallback rate will feel higher than the headline 5%, and it is worth logging stop_reason to see how often it fires on your own traffic. Both models also carry 30-day mandatory data retention with no zero-data-retention option - a constraint in its own right, separate from the classifier question.

Questions people ask

What is the difference between Fable 5 and Mythos 5? They are the same underlying model. Fable 5 ships with safety classifiers and is generally available; Mythos 5 has those classifiers removed and is restricted to organisations approved under Project Glasswing. Same capability, different access policy.

What does "Mythos-class" mean? It is Anthropic's name for a new capability tier positioned above the Opus class. Fable 5 is the first model in it, and Opus 4.8 becomes the fallback model beneath it.

What do Fable 5's safety classifiers do? They cover cybersecurity, biology and chemistry, and suspected distillation. When one fires, the query is answered by Opus 4.8 instead and the API returns stop_reason: "refusal" as an HTTP 200 - Anthropic says this happens in fewer than 5% of sessions.

Who can use Mythos 5? Only organisations approved under Project Glasswing - Anthropic's restricted programme, run with the US government, for cyberdefence, infrastructure and biology-research organisations. Access is arranged through an Anthropic, AWS or Google Cloud account team.

What do Fable 5 and Mythos 5 cost, and how large is the context window? Both are priced at $10 per million input tokens and $50 per million output tokens, with a 1,000,000-token context window and up to 128,000 tokens of output per request.


Sources & caveats. The tier definition, the "same model, different access policy" statement, the classifier and refusal mechanics, the under-5% figure, the specifications and pricing all come from Anthropic's 9 June 2026 announcement and its live Platform documentation, checked 18 July 2026. The two "safeguards removed" benchmark rows are vendor-reported figures relayed by a single secondary aggregator and are included only to illustrate what the classifiers gate; they are not independently verified.