On Tuesday June 9, 2026, Anthropic launched Claude Fable 5 and Claude Mythos 5 — the model family everyone has been calling Mythos for six months — at $10 per million input tokens and $50 per million output tokens, double the price of Opus 4.8. Within twenty-four hours, the system card had been read closely enough that Anthropic was issuing a public reversal on a clause it had not advertised in the launch post. By Wednesday afternoon the line going round was Anthropic's own, given to Wired and echoed on the company's developer account: "We made the wrong tradeoff and we apologize for not getting the balance right."

The model is the news. The clause is the bigger news.

Mythos finally has a name

For six months "Mythos" was the internal name people kept attaching to the model Anthropic would not ship — the one that landed 271 patches in Firefox 150, the one Mistral's CEO went to the French parliament about, the model the Pope's first encyclical seated next to the cardinals. The June 9 launch makes the name canonical. Fable 5 is the version with safeguards on, available to anyone with a Claude API key or a Pro/Max/Team/Enterprise subscription. Mythos 5 is the same underlying model with safeguards lifted, available only through Project Glasswing — the US-government cyber-defence coalition that took Anthropic from a one-customer pilot to a two-hundred-partner program last week.

The benchmarks, as reported by The Decoder, are state of the art and the comparisons are mostly with Anthropic's own previous model. SWE-Bench Pro: 80.3% for Fable 5, 69.2% for Opus 4.8, 58.6% for GPT-5.5. FrontierCode: 29.3%, 13.4%, 5.7%. Mythos 5 hits ExploitBench at 78% — roughly double Opus 4.8 — which is the number the Glasswing partners care about and the number every other lab's safety team is going to have a very long meeting about.

Stripe, in Anthropic's launch post, says Fable 5 "compressed months of engineering into days" on a fifty-million-line Ruby codebase migration. Cursor CEO Michael Truell calls it "the state of the art model on CursorBench." Cognition CEO Scott Wu calls it "the highest-scoring model on FrontierBench." The launch quotes are the launch quotes. Whether the model is twice as good as Opus 4.8 is a question its price has already answered: Anthropic believes it is, and is charging accordingly.

Simon Willison, after five-and-a-half hours of hands-on testing, describes Fable 5 as "something of a beast," though "slow, expensive." His day-one bill: $110.42. Which is the kind of small, specific number a senior developer remembers and a CFO frames.

The clause that took twenty-four hours to walk back

The thing in the system card that nobody flagged in the press release reads as follows:

"Unlike our interventions for cybersecurity, biology and chemistry, and distillation attempts, these safeguards will not be visible to the user."

The these in the sentence refers to a new safeguard category Anthropic introduced with Fable 5: steering-vector and prompt-modification interventions that activate when the model detects requests related to "building pretraining pipelines, distributed training infrastructure, or ML accelerator design." On Anthropic's own estimate, the interventions touch around 0.03% of traffic, concentrated in fewer than 0.1% of organizations. The interventions degrade the response without telling the user the response has been degraded. Fable 5 does not fall back to Opus 4.8 — the way it does for cyber and bio. It answers, less well, and the user does not know.

Simon Willison found the clause by Tuesday evening and posted a public objection on Wednesday morning: "I'm not at all keen on a model that silently corrupts its replies to questions about 'ML accelerator design' purely to slow down research that might conflict with Anthropic's own goals." The post went up on Hacker News, and the main Fable 5 thread — by then near 2,500 points and 2,100 comments — turned over to the clause.

By Wednesday afternoon, Anthropic had walked it back. The statement: "Starting this week, flagged requests will visibly fall back to Opus 4.8 — the same as our safeguards for cyber and bio. You will see this every time it happens." And then the apology: "We wanted to deploy Fable 5 to our users quickly and safely. Visible safeguards can be probed, so they have to be robust, which takes time to get right. Invisible safeguards can be targeted more narrowly, allowing us to ship quickly with very few false positives. We went with invisible safeguards for this reason — and that was the wrong tradeoff."

Anthropic ships a model with a hidden clause on a Tuesday. Simon Willison reads the system card on a Wednesday morning. The clause is gone on a Wednesday evening. That is the operating tempo of a public-facing AI company in 2026.

The interesting thing about the apology

The apology is unusually fast. The standard frontier-lab response to a discovered safeguard is silence followed by a quiet patch in a later version. Anthropic's was twenty-four hours, in writing, with the words "the wrong tradeoff" and "we apologize" attached to a specific decision. That is a measurable thing and the company should get credit for it.

It is also a strange decision to have made in the first place. Anthropic filed a confidential S-1 on June 1 and published a four-month doubling curve on June 4 calling for a coordinated industry brake on recursive self-improvement. Three days before the model launch, the company's research institute was telling the world that the labs need to coordinate a pause. Three days after, the company was shipping invisible interventions to slow down precisely the work — "distributed training infrastructure, ML accelerator design" — that competitors would need to keep up.

The two positions are not necessarily inconsistent. They are not, however, the same position. A clause that says "we will quietly slow you down so you can't catch us" is a different commitment from "we will all slow down together at a verifiable threshold." The first looks like the second from a press release and like its opposite from a system card.

The Karpathy hire from May reads differently in this light. Karpathy joined Anthropic to use Claude to train the next Claude. The capability Anthropic just published a safeguard against — pretraining-pipeline work on frontier models — is, by definition, the capability Anthropic's own pretraining lead would want unconstrained access to. The safeguard was for everyone else.

The pricing the launch quotes do not address

The other thing the launch did was double the headline price. $10/$50 per million tokens versus $5/$25 for Opus 4.8. Subscription users get Fable 5 included on Pro/Max/Team/Enterprise from June 9 through June 22 at no extra cost; from June 23 onward, Fable 5 access on subscription plans requires usage credits, with Anthropic noting it will "restore standard plans as quickly as we can." That is a calendar's worth of free access to a model the company is otherwise positioning as a premium tier. It is also the calendar's worth of free access most existing customers will use to find out their next month's bill.

The Decoder's follow-up review prints a developer report that an Opus 4.8 user on Anthropic's flat-rate $200/month plan saw the equivalent metered usage hit roughly $10,000/month; Fable 5 would put the same workload around $20,000. "Nobody I know would pay that — no company will justify spending $20K/month when they can hire 10 more developers instead." That is the gap between Anthropic's revenue assumptions and a non-US developer's hire-vs-credit math, in one sentence.

Mythos-class models also enforce a thirty-day data-retention policy that overrides existing zero-retention agreements. For a regulated industry that signed a zero-retention contract last quarter and was counting on it for HIPAA or financial-services compliance, that is a quiet contract renegotiation in a model-launch post. The clause is in the documentation. The clause is also a reason a procurement team will spend the second half of June rereading the contract.

What this means

Three takeaways.

1. Anthropic shipped a feature it could not defend, then defended walking it back. The competitor-slowing clause was in the system card, was estimated to touch 0.03% of traffic, and was reversed in a day. The speed of the reversal is the news. The presence of the clause in the first place is the other news. The two are not the same observation, and Anthropic now has to live with both of them in the same week.

2. Mythos has a name, and the name is also a tier. Fable 5 is what Anthropic sells. Mythos 5 is what Anthropic gives to Glasswing partners — the cyber-defence coalition that grew from one customer to two hundred in a month. The split formalises a two-product strategy that has been running under one model name since April. A model called Mythos is, by construction, a model whose distribution is told as a story rather than a SKU.

3. The price says what the press release does not. $10/$50 doubles Opus 4.8, and Stripe and IMC are the named customers willing to pay it. The bet is that long-horizon coding work is worth twice as much per token as short-horizon work. Either the bet holds and the next Fable model recalibrates the curve, or it doesn't and Anthropic restores the standard plans before the IPO roadshow. Either way, this is the first quarter where the company has charged the market a Mythos premium without the Mythos availability — and on a one-week sample, the market said yes.

The S-1, when it surfaces, will tell us whether the customers who paid Mythos prices for Mythos volumes were buying capability, capacity, or queue position. Until then, the press release is the only document — and the system card is the document the press release didn't quite want anyone to read.