Anthropic released Claude Fable 5 on June 9, its first publicly available Mythos-class model, pricing it at $10/$50 per million tokens. Benchmarks are strong. The developer reaction is split between genuine awe at hard tasks and cost shock when the invoice arrives.

Anthropic released Claude Fable 5 on June 9 — a Monday, as has become customary — and within 24 hours the developer community had collectively run up tabs that, in at least one documented case, cleared $110 before lunch.

Fable 5 is Anthropic's first publicly available "Mythos-class" model, a new tier the San Francisco company positioned above its existing Opus line. The specs are large: a 1-million-token context window with up to 128,000 output tokens per request. The price is larger. At $10 per million input tokens and $50 per million output tokens, it costs roughly twice what Opus 4.8 ran and about three times Sonnet — though Anthropic notes that's less than half the price of the previous Mythos Preview access tier. Anthropic simultaneously announced Claude Mythos 5, the same underlying model with certain safety guardrails lifted, initially deployed through Project Glasswing for U.S. government cybersecurity work.

On benchmarks, the lead is substantial and third-party corroborated. Fable 5 scores 80.3% on SWE-Bench Pro; Opus 4.8 sits at 69.2%, GPT-5.5 at 58.6%, Gemini 3.1 Pro at 54.2%. Stripe reported in early testing that the model compressed what would have been a two-month engineering project — a codebase-wide migration across 50 million lines of Ruby — into a single day.

The developer reaction tracked roughly along task difficulty. Simon Willison, whose hands-on verdicts are treated as a reliable early signal in the builder community, called Fable 5 "something of a beast" and documented it finding bugs in his own library while solving the problem he'd assigned it. Jamie Marsland of Automattic ran a single-shot WordPress block theme build from a screenshot and declared it "next level." On Hacker News, one developer working on database migrations reported a 46x reduction in memory allocations. Another, working on CRDTs, said the model derived correct invariants and wrote verification fuzzers with minimal prompting. At least one commenter said Fable felt "like it's coming for my job" in a way earlier models hadn't.

The enthusiasm cooled against two friction points. The first is cost in practice: Willison's $110 day was widely shared as a data point, not a warning — but one Hacker News user reported a single code-review prompt burning $92 across subagents without completing. Subscribers on the highest reasoning tiers described exhausting five-hour usage windows in eight minutes. The consensus that emerged is that the premium is defensible for hard, open-ended work and indefensible when you point it at tasks Opus 4.8 would have finished.

The second friction point is the safety fallback. Anthropic ships Fable 5 with a classifier that routes queries it flags as cybersecurity, biology, or chemistry risks to Opus 4.8 instead, notifying the user. Anthropic says this fires in fewer than 5% of sessions. Developers who triggered it say the rate felt higher: medical imaging pipelines, laboratory automation scripts, and at least one music firmware project were all flagged and downgraded. The Register ran a piece titled "It blocked us at 'hello.'" Nathan Lambert's critique, the most-shared dissent of the launch week, called the approach "safety fables" and argued that "an AI model that gets less intelligent automatically without notifying me is categorically misaligned AI." His sharper claim — that the selective enforcement around AI research and distillation reads more like competitive protection than safety — remains contested but has not been substantively answered by Anthropic.

A third story emerged quieter: a June 10 Hacker News thread about AWS Bedrock requiring customers to share data with Anthropic to access Mythos-tier models drew 416 points and 251 comments — the most-discussed Anthropic story of the week, and one neither the launch announcement nor most coverage addressed.

Fable 5 is the latest beat in a 2026 release cadence that has included Cowork in January, Opus 4.6 in early February, Sonnet 4.6 two weeks later, and a string of tooling releases through March. The "wake up babe, Claude dropped again" meme is no longer a joke; it is a documentation problem. The thing to watch now: whether the safety classifier false-positive rate actually comes down as Anthropic has promised, what the AWS data-sharing terms actually say, and whether Fable's pricing holds once OpenAI and Google finish calibrating their own responses.