Anthropic released Claude Fable 5 on June 9, 2026 — the first time its top-tier "Mythos-class" technology has been made generally available to the public. For developers, three things matter immediately: it is the most capable model Anthropic has ever shipped to everyone, it is free to try for roughly two weeks, and after that window it costs twice as much as Opus 4.8 to run. Every number below is traced to Anthropic's own announcement and documentation.
What "Mythos-class" actually means
Anthropic shipped two models on the same day. Claude Fable 5 (API ID claude-fable-5) is the version with safety classifiers enabled, and it is the one anyone can use. Claude Mythos 5 (claude-mythos-5) is, in Anthropic's words, "the same underlying model" with those safeguards lifted in certain areas — and it is not generally available. Mythos 5 is restricted to approved customers in Anthropic's "Project Glasswing," which the docs describe as limited to partners and select biology researchers. So when commentators say "Mythos is finally public," what they mean precisely is that the safeguarded sibling, Fable 5, is public. The unguarded Mythos 5 is not.
Because Fable 5 carries those classifiers, it can decline some requests. Anthropic built that into the API: when Fable 5 refuses, the Messages API returns a successful HTTP 200 response with stop_reason: "refusal" and reports which classifier triggered — it is not an error. A separate fallbacks parameter lets the request automatically retry on another Claude model, and you are not billed for a refusal that produced no output. This is a meaningfully different integration model from previous Claude releases, and worth handling explicitly if you build on Fable 5.
The benchmark numbers
Anthropic positions Fable 5 as state-of-the-art on nearly every capability benchmark it tested. The headline coding result is SWE-bench Pro, a harder variant of the standard software-engineering benchmark: Fable 5 scores roughly 80%, against 58.6% for OpenAI's GPT-5.5 and 54.2% for Google's Gemini 3.1 Pro. That is a lead of more than 21 percentage points over GPT-5.5 on the same test, and it clears Anthropic's own Opus 4.8 (69.2%) by about 11 points. On the more familiar SWE-bench Verified, Fable 5 lands around 95%. And on OSWorld-Verified — a live computer-use benchmark run on a real Ubuntu desktop — the launch chart lists 85.0%.
As always, these are vendor-reported figures and should be read with the usual caution. But the throughline is clear: the largest gains are in long-horizon, agentic work — software engineering and computer use — rather than single-shot question answering.
The Pokémon demo that makes the point
The capability claim Anthropic chose to dramatize was a game. Fable 5 played Nintendo's Pokémon FireRed "from start to finish using only raw game screenshots — with no maps, navigation aids, or extra game-state information." The detail that matters is the comparison: earlier Claude models could not beat FireRed even when developers handed them a "complex helper harness" of extra tools. Fable 5 finished it with what Anthropic calls a "minimal, vision-only harness." It is a vivid way of showing that the model can perceive an unfamiliar environment and pursue a goal across thousands of steps without hand-holding — the same property that drives the agentic benchmark scores. (Anthropic's other launch demos included autonomously playing Factorio, deriving planetary orbits from first principles to predict a solar eclipse, and coding a fluid simulation set to music.)
Pricing, the free window, and the catch
Fable 5 has a 1M-token context window and can produce up to 128k output tokens per request. It is priced at $10 per million input tokens and $50 per million output tokens — exactly double Opus 4.8's $5 / $25. That makes it the most expensive standard Claude model to run per token.
The offset is a free introductory window. From June 9 through June 22, 2026, Fable 5 is included at no extra cost for Claude Pro, Max, Team, and seat-based Enterprise subscribers. Starting June 23, usage moves to credits at the $10 / $50 rate. If you want to evaluate whether Fable 5 is worth the premium on your own workloads, the unmetered window is the time to do it — after that, the cost model changes.
How "thinking" works on Fable 5
One practical change for builders: Fable 5 uses what Anthropic calls "adaptive thinking," and it is always on. You cannot disable thinking, and the raw chain-of-thought is never returned (you can request a summarized version). Instead of a manual thinking-token budget, you control depth with the effort parameter — low, medium, high (the default), xhigh, and max. Higher effort means the model spends more tokens reasoning, calling tools, and explaining; lower effort trades some capability for speed and cost. In Claude Code, the default is high, and the "ultracode" option in the effort menu pairs xhigh effort with permission to launch multi-agent workflows — it is not a secret sixth tier. We cover when to reach for max thinking, and what it costs, in a companion guide.
What this means for vybecoding users
If you build with Claude here, the practical takeaways are: try Fable 5 before June 22 while it is free, because after that it is a deliberate $10 / $50 decision rather than a default; expect its biggest wins on agentic, multi-step coding and computer-use tasks; and budget for the refusal-and-fallback behavior if you wire it into production. It is a real step up in capability — with a real price tag attached.
Sources
Written by the vybecoding.ai editorial team
Published on June 10, 2026