Claude Fable 5 Is the Public Half of a Model Anthropic Won't Fully Release

Fable 5 tops Anthropic's own coding benchmark by 11 points and ships to every paid plan. Its twin, built from identical weights, is restricted to about 50 organizations hunting zero-days that survived 27 years in critical software.

On June 9, 2026, Anthropic shipped Claude Fable 5 to the Claude API and every paid plan. It scores 80.3% on SWE-Bench Pro, Anthropic's agentic-coding benchmark — 11.1 points ahead of Claude Opus 4.8's 69.2%, and a bigger single-generation jump than the company has published on that test before. On FrontierCode, Fable 5 more than doubles Opus 4.8's score, 29.3% versus 13.4%. But Fable 5 is not the most capable model Anthropic built this quarter. That model is Claude Mythos 5 — built from the same weights — and it isn't for sale.

Same weights, two products

Fable 5 and Mythos 5 share an identical base model. The difference is a safeguard layer sitting in front of Fable 5's public deployment: requests touching offensive cybersecurity tactics, biology or chemistry hazard research, or attempts at model distillation get automatically routed to the more conservative Claude Opus 4.8 instead. Anthropic has acknowledged this layer "can block harmless requests" too — a false-positive cost the company is evidently willing to pay. Fable 5 prices at $10 per million input tokens and $50 per million output tokens, double Opus 4.8's rate.

Mythos 5 removes those guardrails and is reserved for a small group of partners. Anthropic has said it doesn't intend to make a Mythos-class model generally available at all — access runs instead through a separate initiative built specifically around what that uncapped capability is for.

What the restricted twin is actually for

That initiative is Project Glasswing, and it reframes why Anthropic split the release in the first place. Using an earlier Mythos preview, Anthropic and its partners found thousands of zero-day vulnerabilities — flaws unknown to the software's own maintainers — across every major operating system and browser, including some that had gone undetected for 16 to 27 years in projects maintained by small volunteer teams. Glasswing's launch partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, with access extended to more than 40 additional organizations that build or maintain critical infrastructure software. Anthropic is backing the effort with up to $100 million in usage credits plus $4 million in direct donations to open-source security groups.

That's the actual shape of the tradeoff: the same model capability that finds a 27-year-old flaw in a browser is the capability that could just as easily find a way to exploit one. Anthropic's answer isn't to withhold the capability — it's to sell a deliberately hobbled version to everyone and hand the unhobbled version only to the defenders who can use it at scale before attackers do.

What this means if you're choosing a coding model today

For anyone deciding between Fable 5 and Opus 4.8 for agentic coding work, the calculus is straightforward: Fable 5 costs twice as much per token, but an 11-point jump on SWE-Bench Pro is a large enough gap to matter for long-horizon agentic runs, where errors compound across many steps. Budget for the higher rate on tasks where correctness is worth it, and keep Opus 4.8 as the default for routine work. And if Fable 5 unexpectedly hands back an Opus-quality answer on something touching security, bio, or chemistry, that's the safeguard layer redirecting the request, not a regression — a useful signal in itself about where the line between "public" and "restricted" AI capability currently sits.