Anthropic Sells Claude Fable 5 at $50 per Million Tokens and Keeps Mythos 5 for Partners

Anthropic released Claude Fable 5 and Claude Mythos 5 on June 9, the first two models the company files under its Mythos class ¹. Both run on one underlying model. Fable 5 carries classifier safeguards for cybersecurity, biology, chemistry and distillation, and each blocked request falls back to Claude Opus 4.8 ¹. Mythos 5 lifts those safeguards in some areas and stays with Project Glasswing partners and a short list of biology researchers ¹. Pricing lands at $10 per million input tokens and $50 per million output tokens, double the Opus 4.8 rate ². Michael Truell, chief executive of Cursor, gave the launch its sharpest customer line. "It's opened up a class of long-horizon problems that were out of reach for earlier models," Truell said ¹.

Why does a price list belong on a ledger about superintelligence? Because the ledger scores public evidence. A model rented at a stated rate is public in a way a twelve-partner preview is. The question is how far Anthropic's own benchmark table carries that evidence. Its answer sets the size of the step.

Anthropic sells Fable 5 and keeps Mythos 5 for partners

Two names, one model. Anthropic's spokesperson put the split in a sentence to Semafor: "The difference [in Fable and Mythos] is not what the model can do, but what our safeguards will allow" ⁶. Read the sentence for its possessive. The capability belongs to the model; the permission belongs to Anthropic. Fable 5 is the permission most people get.

Three classifiers sit in front of Fable 5, and each one hands a flagged request to Opus 4.8 ¹. Anthropic says the safeguards fire in under 5 percent of sessions on average ¹. TechCrunch quotes the company's own test figure of "at least 95% of Fable sessions running entirely on the model's own responses" ². An external bug bounty, in Anthropic's words, "produced no universal jailbreaks in over 1,000 hours of testing" ². Every Mythos-class request is kept for 30 days, used for safety work and then deleted ¹.

Mythos 5 goes where Mythos Preview went. Project Glasswing opened in April with twelve founding partners, among them Amazon Web Services, Apple, Google, Microsoft, NVIDIA and JPMorganChase ⁷. More than 40 further organizations got access for defensive security work ⁷. Anthropic committed $100 million in usage credits and $4 million in donations to open-source security groups ⁷. It reports thousands of high-severity vulnerabilities found, including a 27-year-old flaw in OpenBSD and a 16-year-old flaw in FFmpeg ⁷. Glasswing's page also states that Anthropic "has also been in ongoing discussions with US government officials about Claude Mythos Preview" ⁷. Preview users move to Mythos 5 at once, and a cybersecurity access program and a biology trusted-access program are planned ¹.

Subscribers get a window. Pro, Max, Team and seat-based Enterprise plans include Fable 5 through June 22 ¹. From June 23 it draws usage credits until capacity allows a return to the standard plans ¹. Simon Willison, who tested the model for five and a half hours on launch day, priced the window from his own account. "The model is available 'until June 22nd' on the subscription plans (I'm on $100/month Max at the moment), after which it will be billed extra," Willison wrote, and his own bill for one day of testing came to $110.42 ³.

The case for the move

Truell's sentence is the strongest case for a step up, because it describes a horizon, and horizon is the thing the definition asks for. Mario Rodriguez, chief product officer at GitHub, said the same thing in more words. "In our early testing, it took on complex, long-horizon coding tasks with a level of autonomy and reliability that exceeded previous benchmarks," Rodriguez said ¹. Scott Wu, chief executive of Cognition, reported it as "the highest-scoring model on FrontierBench, Cognition's frontier coding eval" ¹.

Stripe supplies the number that carries the most weight. Anthropic reports that Fable 5 migrated a 50-million-line Ruby codebase in one day, work Anthropic says a team had scoped at two months ¹. A one-day migration by a public model at a public price is a different kind of claim from a benchmark score. It names a task, a size, a duration and a customer who can be asked about it.

Mythos 5's numbers reach further. In drug design, Anthropic reports protein design tasks running about ten times faster than human operators, with strong candidates on nine of fourteen targets ¹. Scientists in blinded comparisons preferred its molecular biology hypotheses about 80 percent of the time against Opus-class models ¹. Willison's own poke at the model added the plumbing figures: a one million token context window, 128,000 maximum output tokens and a January 2026 knowledge cutoff ³.

The case for a smaller step

Three reasons shrink the step.

First, every benchmark in the launch note is Anthropic's. The definition accepts a vendor harness only when an independent board reproduces it, and on June 9 the reproductions were a day or more away. Willison marked the limit of what an outsider could see. "Anthropic haven't said anything about model size, so all we have are tea-leaves, but the speed, pricing and my own poking at its knowledge make me think that it's a large model," he wrote ³. He also found the upgrade guide "substantially thinner than the similar guide for Opus 4.8" ³. A large model at double the price, described by a thin guide, is a product. Evidence is a measured result on an uncontaminated test, and that arrives later.

Second, the stronger tier stays private. Mythos 5 reaches Glasswing partners and select biology researchers, and the drug design and hypothesis figures belong to it ¹. On this ledger a system only partners can run earns a reported step at most, whatever its numbers say. Richard Whaling, lead researcher at Charlemagne Labs, offered Fortune a second reading of the restriction in April. "I share Anthropic's concerns around Mythos's potential misuse, but I think there is also a resource limitation at play," Whaling said ⁵. His tell is the word "also": he accepts the safety case and adds a GPU budget beside it. A model withheld for capacity reasons is a smaller capability signal than a model withheld for danger.

Third, the safeguards reached further than the launch note said. The researchers Anthropic sold the model to found out first. Fortune reported on June 10 that one paragraph in the 319-page system card described limits on frontier AI research work, limits the card called "not visible to the user" ⁴. Jeremy Howard, head of Fast AI, took the release strategy apart on its own terms. "Anthropic has chosen the opposite of the safe path: they are allowing themselves, the current top lab, to use their top model for frontier AI research," Howard said ⁴. Nathan Lambert, an open-model researcher at AI2, said this: "To have my access to the cutting edge models for my work rug pulled in an under the table fashion is appalling" ⁴. Anthropic reversed within a day. "We made the wrong tradeoff, and we apologize for not getting the balance right," a spokesperson told Fortune ⁴. So Anthropic, which priced its model on its own benchmark table, spent day two apologizing for what the table left out. Howard's objection stands after the apology. The visible version of the limit still routes frontier research work away from the top model for everyone except Anthropic.

Where this sits on the ledger

The reading measures four clauses: unsupervised expert work across occupations, ten gigawatts acting as one machine, measured self-improvement, and third-party proof of execution. This launch touches the first clause through the Stripe migration and the customer reports of long-horizon coding. It touches the third clause in an awkward way. Anthropic reserves the strongest research use of its top model for itself, which is the self-improvement clause in practice. Howard's complaint is that the reservation is the point. The compute clause sits untouched. Verification moves the wrong way: the one outside check on the table so far is Willison's five and a half hours and a bill.

Capability moves 0.7, at reported confidence. A vendor announcement of a public product sits in the reported band. The Stripe figure and the customer quotes lift it toward the top of that band. Vendor-only benchmarks and the private Mythos tier hold it under confirmed. A ledger that moved a full point on every launch note would have crossed fifty by now.

By the numbers

• $10 per million input tokens and $50 per million output tokens, double the Opus 4.8 rate ²
• 50 million lines of Ruby migrated in one day at Stripe, against a two-month team estimate ¹
• Under 5 percent of sessions trigger a safeguard, by Anthropic's own average ¹
• Nine of fourteen protein targets produced strong drug design candidates with Mythos 5 ¹
• 1,000 hours of external bug bounty testing, in Anthropic's account ²
• Twelve founding Glasswing partners, more than 40 further organizations, and $100 million in credits ⁷
• 319 pages in the Fable 5 system card, with the frontier-research limit in one paragraph ⁴
• $110.42 for one day of Willison's testing on a $100-a-month plan ³

What to watch

An independent board reproducing the Terminal-Bench and coding figures on its own harness would turn this reported step into a confirmed one. Stripe describing the Ruby migration in its own words, with the error rate and the human review it took, would do the same for the autonomy clause. The cybersecurity access program and the biology trusted-access program are the markers for Mythos 5: a stated date, a stated number of organizations, and any government condition attached to them. A reversal would arrive as a retracted benchmark, a harness dispute, or a regulator closing the door Anthropic opened on June 9.