Anthropic has released Claude Fable 5. It's the first model in the new "Mythos" class — a tier above Opus — made safe enough for general use. Alongside it comes Claude Mythos 5: the same base model with its cyber safeguards lifted, restricted to a small set of government and infrastructure partners.
The benchmark numbers are real, and they're a jump. SWE-Bench Pro: 80.3% versus 69.2% for Opus 4.8 and 58.6% for GPT-5.5. On Cognition's FrontierCode it more than doubles Opus. Stripe ran an early test and finished a migration across a 50-million-line Ruby codebase in a day. By hand, that was two months of team work.
If you build SaaS, the headline isn't "new best model." It's two structural shifts you now have to design around.
The frontier moved up, and got more expensive
Fable 5 costs $10 per million input tokens and $50 per million output. That's twice Opus 4.8 ($5 / $25). Anthropic priced it as "less than half" of Mythos Preview, so by frontier standards it's a discount. For your cost model, it's still the most expensive Claude you can call.
The catch worth knowing: it's also more token-efficient. The number that matters for your bill isn't price per token, it's cost per task. A model that's twice the price but finishes in half the turns can land in the same place. You only learn that by measuring, not assuming.
If you run a credit-based model, this is the moment to stress-test your pricing. At $50 per million output tokens, an end user who picks the top model and runs long agentic tasks can burn your margin fast. Per-model pricing, token tracking, and budget caps stop being nice-to-haves.
Anthropic just shipped a fallback, the pattern you already need
Here's the part that's easy to miss. Fable 5 routes flagged prompts — cybersecurity, bio and chemistry, distillation — to Opus 4.8 instead of refusing them. Fewer than 5% of sessions hit it. In plain terms: the market leader built an automatic model fallback into its flagship.
That's the same pattern a serious multi-provider setup gives you out of the box: route to a different model when the first one can't or shouldn't answer. If you've been treating provider fallback as over-engineering, this is the argument that it's table stakes.
What I'd do this week
Add claude-fable-5 to your model catalog, as an opt-in "max capability" tier, not the default. The price makes it a deliberate choice, not a silent one.
Map the credit cost per model. Run one real task on Fable versus Opus and compare cost per task, not per token.
Mind the date. Through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans. From June 23 it needs usage credits. If your team leans on the subscription, plan for the cliff.
A model this capable is good news. The work it shifts onto you is pricing discipline and routing. Both are architecture, not afterthoughts.
Sources: Anthropic, THE DECODER