Fable 5 vs Mythos 5 — one model, two names.

The names are the same word in two languages. Fabula is Latin, mythos is Greek — both mean "that which is told." That's the architecture stated as etymology: one story, two tellings. Anthropic's own framing is that Fable 5 is "a Mythos-class model made safe for general use," and Mythos 5 is the same model with safeguards lifted in some areas.

The split is the product decision. In March, Anthropic disclosed a model this capable and said it wouldn't ship. In June it shipped twice — the raw capability for the vetted, the classified version for everyone with an API key. Not a big model and a small model, not a pro tier and a cheap tier. One model, two trust postures. Both priced at $10/$50 per million tokens — the gate is vetting, not money.

Today: Project Glasswing partners, with the cyber safeguards lifted. Next: select biology researchers, with bio/chem safeguards lifted but cyber intact. After that: a broader trusted-access program, planned but not detailed.

Read the sequence carefully — access isn't all-or-nothing. The biology expansion lifts a different safeguard for a different audience; a vetted virologist gets the bio capability and still hits the cyber classifier. Anthropic is unbundling the model's danger surface and granting it back one domain at a time, per audience. The early receipts from that gated lane are in the announcement: Mythos 5 ran autonomous genomics research across 138 animal species, and scientists preferred its hypotheses ~80% of the time in blinded comparisons against Opus-class models — vendor-curated numbers, flagged as such.

For everyone else, Fable 5 is the access program. Available everywhere June 9: full availability on the API and consumption-based Enterprise, included in paid-plan limits until June 22, usage credits after. The clock math lives on the hub.

1 — Cybersecurity. Blocks queries on some exploitation and offensive-cyber topics. This is the classifier with receipts attached: external testing found zero harmful single-turn cyberattack-planning requests succeeded, and 1,000+ hours of external red-teaming found no universal jailbreaks. Those are vendor-reported security results — the same Ch 24 rule applies to safety numbers as to capability numbers: a signal, not a receipt.

2 — Biology/chemistry. Blocks most bio/chem requests on dual-use concern. The dual-use example is stated plainly in the announcement: a model that can design a gene therapy can design a dangerous virus. The separate receipt: Anthropic's internal protein experts report it accelerated their drug-design work by around 10x. The classifier doesn't distinguish your intent; it gates the domain.

3 — Distillation. Flags attempts to extract Claude's capabilities for competing models, particularly in authoritarian countries. The least discussed of the three and the most telling — it treats the model's own behavior as the asset being stolen, not just the harm being caused.

Trip a classifier and what happens depends on the surface. In the Claude apps and Managed Agents, fallback is built in — the response comes from Claude Opus 4.8 and you're told it happened. On the raw Messages API, a blocked request returns an error rather than a model response, and you aren't charged for it; fallback to Opus 4.8 is an opt-in, billed at Opus pricing, or you roll your own with the SDK's support for exactly this. The base rate matters more than the mechanism: more than 95% of Fable sessions involve no fallback at all. If your sessions are in the other slice, that's routing information, not bad luck — see the operator move below.

The data terms ride along with the safeguards: Fable 5 requires a limited 30-day retention window — retained data is used only to detect and prevent serious misuse, never to train Claude — and admins must accept updated terms in the Claude Console before the model works. Human access to retained data is logged; deletion happens after 30 days in almost all cases. And on alignment, Anthropic reports Mythos 5's misaligned-behavior level as low and similar to Opus 4.8 — Fable 5 similar, same underlying model.

Anthropic's table footnote is the most operator-relevant sentence in the launch: reported scores are within 1–3 percentage points between Mythos 5 and Fable 5, and the table shows the higher of the two. On most rows that's a rounding argument. On the starred rows it isn't — those show a larger gap because the blocking safeguards trigger on cyber- and bio-related questions, and Fable 5 performs closer to Opus 4.8 there due to fallbacks.

* "Higher of the two" — on starred rows, the published number leans toward the gated model. The full table — thirteen benchmarks, including the starred Terminal-Bench 2.1 at 88.0% — is on the benchmarks page.

So read ExploitBench honestly: 78.0% is in practice the score of a model you can't buy. The model you can buy sits closer to Opus 4.8's 40.0% on that work — by design, that's the safeguard functioning. A starred launch number is partly an ad for the gated twin. The fix is the book's standing one (Ch 25): your private eval runs against claude-fable-5 with the classifiers on, which makes it the only benchmark measuring the product you'll pay for.

• March 2026 — a Fortune leak discloses an internal model code-named Mythos; formal references follow in Anthropic's safety materials. Disclosed Preview numbers: SWE-bench Verified 93.9%, SWE-bench Pro 77.8%, OSWorld 81%.
• Then — Anthropic states Mythos Preview will not be made generally available. Project Glasswing ships instead as the operator-facing successor.
• May 6, 2026 — this site's research note reads the disclosure as a capability ceiling Anthropic won't productize: "Mythos isn't coming." Wrong call, kept on the record.
• June 9, 2026 — it comes — split in two. Mythos 5 stays gated, Fable 5 goes on sale at less than half the price Mythos Preview cost its partners.

The mechanism behind the wrong call is worth keeping: "disclosed but withheld" wasn't a refusal, it was staging. Anthropic didn't choose between shipping and withholding — they built the architecture that lets them do both. Expect the next frontier release to wear the same shape: raw model gated, classified twin sold, fallback to the previous Opus.

Here's the move the split hands you. Fable 5 costs $10/$50 — exactly 2× Opus 4.8's $5/$25. When a classifier trips, you get an unbilled error on the API or an Opus 4.8 answer in the apps — either way, the frontier delta you came for is gone mid-task. So if your workload is cyber- or bio-adjacent — security research, exploit analysis, anything in the domains the starred rows flag — the cost isn't double-billing, it's disruption. Route it to Opus 4.8 directly and keep the 2× for work that stays out of classifier territory.

This is Ch 9's discipline pointed at model choice: map your blast radius before the incident, not during. Know which of your workloads sit near the three classifiers before a mid-task fallback maps it for you. Everything else — the 95%-plus of sessions that never trip a classifier — goes to Fable 5 during the June 9–22 plan window, where the eval is free and the transcripts are yours (Ch 25). The tier list stays the live routing table.