Anthropic has introduced Claude Fable 5, its most capable model to date and the first entry in a new model tier that sits above the Opus line. Fable 5 ships with a 1-million-token context window and 128K maximum output, and is priced at $10 per million input tokens and $50 per million output tokens, roughly double Opus 4.8's $5/$25.
A deliberate step up, not an Opus refresh
Anthropic is keeping Opus 4.8 as the workhorse of the lineup and positioning Fable 5 above it as the model to reach for when raw intelligence matters more than cost: long-horizon agentic runs, hard reasoning, and knowledge work where a better answer pays for itself. The 1M context window carries no long-context premium, so the headline price is the only number teams need to reason about.
The price is the point
At $10/$50, Fable 5 is the most expensive model Anthropic has shipped, and the gap to Opus 4.8 is a clean 2x on both input and output. That is a tiering decision as much as a compute one: Anthropic is signaling that the frontier of capability now comes with a frontier price, and that most workloads should still run on Opus or Sonnet. It lands in the middle of the broader pricing reckoning that GitHub Copilot's move to per-token billing kicked off. A flagship at $50 per million output tokens is the clearest sign yet that the subsidy era is over, and that model selection is becoming a cost discipline, not just a quality one.
Nearly a drop-in for Opus 4.7 and 4.8
For anyone already on Opus 4.7 or 4.8, Fable 5 is close to a drop-in. The breaking changes are inherited rather than new: fixed-budget extended thinking and the sampling parameters (temperature, top_p, top_k) were already removed on the 4.7 family. The one fresh trap is that an explicit thinking: {type: "disabled"} now returns a 400, so code that simply omits the thinking field keeps working. The minimum cacheable prefix is also smaller than Opus (2048 versus 4096 tokens), a minor win for prompt caching on shorter shared prefixes.
What we'd tell technical leaders
Resist routing everything to the newest, most powerful model by reflex. Fable 5 earns its price on a specific class of hard problems; the rest of your traffic almost certainly does not need it. The move this week is a tiered routing strategy, Fable 5 for genuinely difficult tasks, Opus 4.8 as the default, and Sonnet for high-volume work, backed by token-level cost observability so you can see what each tier actually buys. A 2x price step is exactly the kind of decision that should be made by a router and a dashboard, not by whichever model string a developer pasted in last.
.jpg)
