Anthropic dropped Claude Fable 5 yesterday. First Mythos-class model available to the public. Same tier that found and exploited vulnerabilities in every major operating system and browser earlier this year. 80.3% on SWE-Bench Pro, 29.3 on FrontierCode Diamond, beat Pokémon FireRed using nothing but raw screenshots.
Most developers won’t need it.
The advantage is conditional
Anthropic’s own words:
The longer and more complex the task, the larger Fable 5’s lead over our other models.
The lead shrinks on short tasks. The kind that make up most development work. TrueFoundry’s analysis confirms it. On shorter, well-scoped tasks, Fable 5 and Opus 4.8 are close. Opus 4.8 is the sensible default for most production traffic.
Most development work is autocomplete a function, explain an error, draft a test, refactor a file. Opus 4.8 at half the price handles that.
The math
Fable 5: $10 per million input tokens, $50 per million output tokens. 2x Opus 4.8.
One agentic call (50k input, 15k output) costs $1.25. Chain 12 of those for a real task, roughly $15. Max reasoning effort on hard FrontierCode tasks, closer to $20.
That’s a Stripe-class migration. 50 million lines of Ruby, refactored in a day. Not the kind of task most people hand to an AI.
Speed
Artificial Analysis ranks Fable 5 #71 out of 152 models. 60.3 tokens per second. Fine for long tasks, slower on quick ones.
The reasoning-effort dial helps. Low effort, fewer tokens, faster answers. But then Fable 5 prices for Sonnet-class output.
Safeguards
Karpathy on launch day: the safeguards are “configured to be a little too trigger happy.” Anthropic says they’ll tune them.
About 5% of sessions fall back to Opus 4.8 right now. Fable 5 prices, sometimes Opus 4.8 answers. Makes sense for security researchers and biologists. Doesn’t help with React.
The trial ends June 22
Fable 5 is free on Pro, Max, Team, and Enterprise plans through June 22. Then it moves to usage-based credits. Anthropic says they’ll restore it to standard subscriptions “as quickly as we can.”
The $20/month Pro plan crowd will stick with Opus 4.8 when the meter starts.
Who it’s actually for
Stripe compressed a two-month migration into a day. IMC aced trading-analysis evaluations across the board. Rakuten says “the extra thinking pays for itself” for autonomous operations. Anthropic’s internal team accelerated drug discovery 10x.
Enterprises running long-horizon, multi-step workflows where the model works for hours without human intervention.
The benchmarks are Anthropic’s
Every number comes from Anthropic’s published evals. 80.3% SWE-Bench Pro, 29.3 FrontierCode, 88.0% Terminal-Bench. Not independent tests. The starred rows for cybersecurity and biology are Mythos 5 scores. Fable 5, with safeguards active, performs closer to Opus 4.8 on those tasks.
Marketing is cleaner than reality.
Where this lands
Fable 5 is a specialist tool. Reach for it when Opus 4.8 has plateaued. Long-running, multi-stage, complex enough that 2x pricing is justified by 10x time savings.
For everything else, Opus 4.8 is half the price, faster on routine work, and doesn’t have safety classifiers tripping over security questions.