Claude Fable 5 launches as public model and restricted Mythos 5

Anthropic released Claude Fable 5 on June 9 2026 as the first public member of its new Mythos class, positioned above Opus 4.8 [Dev.to]. The launch bundles two postures of the same underlying model: the publicly accessible Fable 5 and the restricted Claude Mythos 5, which is available only to vetted cyber‑defenders and infrastructure providers through the Project Glasswing program in partnership with the U.S. government [Dev.to].

Fable 5 runs three safety classifiers that monitor requests for offensive cybersecurity content, weaponizable biology or chemistry, and model distillation. When a classifier fires, the API returns a JSON payload with stop_reason: "refusal" and a 200 status, then forwards the request to Opus 4.8, the previous public top‑tier model [Dev.to]. Refused calls are not billed because no output is generated. Pricing is $10 per million input tokens and $50 per million output tokens—exactly double Opus 4.8’s $5/$25 rates [Dev.to]. From launch through June 22, Fable 5 is included at no extra cost on Pro, Max, Team, and Enterprise plans; after that date usage draws from plan credits [Dev.to].

Performance benchmarks show a clear gap on agentic coding tasks. On SWE‑bench Pro, Fable 5 scores 80.3 %, versus Opus 4.8’s 69.2 %, GPT‑5.5’s 58.6 %, and Gemini 3.1 Pro’s 54.2 % [Dev.to]. On SWE‑bench Verified, Fable 5 reaches 95.0 % and Mythos 5 95.5 %, the half‑point difference attributed to the safety fallback [Dev.to]. Anthropic also reports first‑above‑90 % scores on Hex’s analytics suite and top marks on Hebbia’s finance benchmark [Dev.to]. A Stripe case study claims a 50‑million‑line Ruby migration completed in one day with Fable 5, versus an estimated two‑month manual effort [Dev.to].

The dual‑posture launch adds integration steps: developers must handle the refusal response, inspect the classifier field, and optionally invoke a fallback model via the fallbacks parameter or client‑side SDK middleware. The double‑price structure raises cost for token‑heavy workloads, though the two‑week free window lets teams test without charge. The performance advantage appears on long‑context coding and finance reasoning; for routine classification or latency‑sensitive tasks, Opus 4.8 or Sonnet 4.6 remain more cost‑effective [Dev.to].

Editor’s take

Anthropic’s launch makes safety a core contract rather than an afterthought. Teams that ignore the classifier will incur hidden costs in retries and token waste. Engineers who automate the refusal‑fallback loop can capture the performance premium without inflating bills.