Claude Fable 5, one model behind two doors

Yesterday, 09 Jun 2026, Anthropic shipped two models that are one model. Claude Fable 5 is generally available from day one: on the Claude API, on Bedrock, Vertex, and Foundry, inside Claude Code, at $10 per million input tokens and $50 per million out. Anthropic’s own sentence is the headline: Fable 5’s capabilities exceed those of any model they have ever made generally available. Claude Mythos 5 is the same underlying model with the safety classifiers removed, and it is not for sale. It goes only to vetted partners through Project Glasswing: cybersecurity teams and selected biomedical researchers, by arrangement, with the lab watching.

A release like this is usually read down the benchmark column, and the column deserves the look. But the benchmarks are not the news. The news is the structure. For the first time, the line between what the frontier can do and what you may do with it is drawn after training, at the door.

What shipped

Both models come from one training run. Both carry a 1M-token context window, 128k output, and adaptive thinking that is always on. What separates them is policy. Fable 5 ships with three classifiers: one for cybersecurity, one for biology and chemistry, one for distillation attempts. They trigger in roughly 5% of sessions, and when they do, the request is not refused into the void; it is handled by Claude Opus 4.8 instead, unbilled. Mythos 5 has those classifiers lifted, per cohort: cybersecurity partners get the cyber safeguards removed, biomedical researchers the biology and chemistry ones. The price of the open door is $10 in and $50 out per million tokens, double Opus 4.8 and less than half what Mythos Preview cost.

The scores, briefly

The column is real. On SWE-Bench Pro, Fable 5 lands 80.3% against Opus 4.8’s 69.2% and GPT-5.5’s 58.6%. On FrontierCode’s Diamond subset, the hardest 50 of the benchmark’s 150 production-codebase tasks, it scores 29.3% where Opus 4.8 managed 13.4% and GPT-5.5 5.7%. Terminal-Bench 2.1: 88.0%. It is the first model past 90% on Anthropic’s internal analytics benchmark, a 10-point jump over Opus 4.8. On a frontier physics research task it reached in 36 hours what GPT-5.5 landed after four days, on a third of the reasoning tokens. The pattern across the table is consistent: the gap to the rest of the field is widest exactly where the work is longest and hardest.

The doors, not the scores

Look closely at Anthropic’s own benchmark table and the structure shows through the numbers. Several rows carry a star: on the starred benchmarks, cybersecurity and biology among them, Fable 5 in public hands scores meaningfully below Mythos 5, because the classifier intercepts the guarded queries and hands them to Opus 4.8. On ExploitBench the underlying model captures 78%; behind the public door, the work runs at Opus-grade capability instead. The capability exists either way. The door decides whether you meet it.

That is the new thing. Until yesterday, what a released frontier model could do was what you got: safety lived in the weights, baked in by the training run, identical for every customer. Now one training run produces one set of capabilities and ships under two policies. Anthropic decoupled what the model can do from what it will do for you, and moved the second question out of the model entirely, into an access tier.

The model did not split in two. The policy did. What the frontier will do for you now depends on who you are to it.

What the curve prices

For an operating business, the most useful chart in the release is not a leaderboard. It is the accuracy-versus-cost curve on FrontierCode. Fable 5 at its lowest effort setting scores 11.5% at about $5.50 per task; Opus 4.8 at its maximum scores 11.4% at about $10.50. The new model’s floor matches the old model’s ceiling at roughly half the per-task cost. And where Opus 4.8’s curve bends flat past its xhigh setting, Fable 5’s keeps climbing, to 31% at around $19 per task. The new curve begins about where the old curve ends.

The right edge matters to labs racing each other. The left edge matters to a business deciding whether an agent that closes the books or audits the supplier thread is worth running every day. That edge just moved down and to the left.

The trust gradient

The second door is worth a moment, even though almost nobody reading this will pass through it. Project Glasswing is Anthropic’s security program with partners of the order of AWS, Google, Microsoft, CrowdStrike, and JPMorganChase; Anthropic credits it with closing more than ten thousand high- or critical-severity vulnerabilities in widely used software. Those partners, plus selected biomedical researchers, get Mythos 5: the same weights with the relevant classifiers off.

Here is the detail that makes the structure legible. In Anthropic’s automated alignment assessment, Mythos 5’s rate of misaligned behavior scores 2.06 on a 1-to-10 scale, statistically level with Opus 4.8’s 2.05 and better than Sonnet 4.6’s 2.81. Removing the classifiers did not change the model’s character. The gate is not there because the unguarded model is misaligned; it is there because two narrow domains, offensive security and the bench end of biology, are dual-use no matter how well behaved the model is. Which means the gate is not a statement about the model at all. It is a statement about the user.

Read forward, this is a quiet repricing of trust. The labs are becoming counterparties that grade you: who you are, what you do, what you can be trusted with determines which capabilities you can buy. Anthropic says it intends to expand trusted access steadily. Expect more of the frontier’s sharpest edge to ship this way, and expect “who are you to the lab” to start mattering to businesses the way a credit rating does.

What to do with this

We wrote in May that the capability overhang is a planning question, and on Saturday that recursion widens it on a clock. Yesterday the clock ticked again, and the structure of the tick is the kindest one possible for an operating business: the gated door is a curiosity, but the open door is the most capable model ever offered to the public, at half the per-task cost of the last generation for matched results, with a million tokens of context at standard rates. Nothing about your eligibility is in question. The only question is whether the model is wired into the actual operation, the actual reservations system, the actual supplier thread, the actual books, or whether it is still a tab somebody opens.

That wiring is the whole of what we do. If the open door has been standing open in front of your operation for a year and nothing has walked through it, start a conversation with us about a Discovery Phase.