Anthropic ships Claude Mythos 5 / Fable 5, a new frontier tier

breakthroughDeveloper

Wednesday, June 10, 2026

Signal

High

Time horizon

This week

Risk

Re-baselining your stack against a model your competitors already have

The News

Anthropic released Claude Fable 5 and the more capable Mythos 5 on June 9 — the same underlying model, shipped in two configurations. Fable 5 is the safeguarded general release, live now on the Claude API and claude.ai at $10 per million input tokens and $50 per million output, and bundled into Pro, Max, Team and Enterprise plans through June 22 before usage credits apply. Mythos 5 is the same model with safeguards lifted in specific domains, gated to Project Glasswing (a US-government collaboration for cyber defenders) with a biology-researcher access program to follow. On Anthropic's own benchmark card the Mythos-class model leads every published category against Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro: SWE-Bench Pro 80.3% (Opus 4.8 69.2%, GPT-5.5 58.6%, Gemini 3.1 Pro 54.2%), the GDPval-AA knowledge-work score 1932 (1890 / 1769 / 1314), Legal Agent Benchmark 13.3% (10.4% / 2.1% / 0.0%), Terminal-Bench 2.1 agentic coding 88.0%, and ExploitBench cybersecurity 78.0% (Opus 40.0%, GPT 34.0%). Anthropic cited Stripe completing a 50-million-line Ruby migration in a day, and protein-design candidates emerging on 9 of 14 targets.

The Read

Strip the model name and read the slope. Anthropic just moved its own frontier roughly a tier above Opus 4.8 — the model most enterprises standardized on this spring — and priced the general version at $10/$50, in the same band as the model it obsoletes. For anyone building on Claude, the practical event isn't "a smarter model exists"; it's that your competitors get the same capability jump on the same day at the same price, so any moat you built on Opus-4.8-class reasoning just reset to zero. The two-tier split is the more strategic move: Fable 5 (safeguarded) is what the market gets; Mythos 5 (unlocked, government-gated) is a clear signal that Anthropic now treats frontier cyber and bio capability as a national-security asset to be metered, not a feature to be shipped. The number that should move your roadmap is the 80.3% on SWE-Bench Pro and 88% on Terminal-Bench: agentic coding has crossed the line where "the model writes and ships the change" is the default assumption, not the demo.

Counterview

Read the benchmark sheet like a prospectus. Anthropic filed its confidential draft S-1 with the SEC on June 1 (TechCrunch), and these are self-reported numbers dropped into an IPO window — "the product is already building the next product" is the dream line to walk into a roadshow with. The footnotes matter: one read of a separate Anthropic capability report argues a chart showing the model out-reasoning human researchers "two times in three" shrinks toward one-in-five once you account for it counting cases where the human had already gone wrong. The capability jump looks real; treat the framing with the skepticism you'd give any company publishing its best numbers the week it courts investors.

Do This Week

Take the single hardest task your current Claude (or GPT/Gemini) tier failed on in the last month and re-run it on Fable 5 before June 22, while it's free in your plan.

Pick one real failure: a multi-file refactor, an agentic migration, a long-context reasoning task you downgraded to human review
Run it on Fable 5 and log whether the failure mode is gone or just moved
Decide one thing by Friday: does this change a build-vs-wait call on your roadmap, or not — and write the reason down either way

For Developer

Take the single hardest task your current Claude (or GPT/Gemini) tier failed on in the last month and re-run it on Fable 5 before June 22, while it's free in your plan.

Pick one real failure: a multi-file refactor, an agentic migration, a long-context reasoning task you downgraded to human review
Run it on Fable 5 and log whether the failure mode is gone or just moved
Decide one thing by Friday: does this change a build-vs-wait call on your roadmap, or not — and write the reason down either way

For Developer