Anthropic shipped Fable 5.1 on 1 September 2026, alongside a restricted-access variant called Mythos 5.1. The model keeps Fable 5’s base pricing but slashes the cost of cache reads by three quarters. For teams running agents in a loop, that shifts the cost of an individual run far more than any benchmark score does.
What changed on 1 September
Fable 5.1 ships with a one-million-token context window by default and a 128,000-token maximum output. Adaptive thinking is always on, and the model is available immediately as claude-fable-5-1 on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Fable 5.1 and Mythos 5.1 refer to the same underlying model: the former is the generally available version with production safeguards, the latter stays open to vetted organisations in cybersecurity and life sciences.
The move echoes what happened across the range, where Claude Sonnet 5’s introductory price became permanent: Anthropic now revises its price sheet almost as often as its capabilities.
Cache reads: the real battleground for agents
An agent working on a codebase, a customer file or a long conversation thread re-reads a stable context on every turn: system instructions, tool definitions, reference files. Without caching, that context is billed at the full input rate on each call. Prompt caching stores those blocks server-side and replays them at a reduced rate, provided the reusable segments are marked.
That is exactly the line item Fable 5.1 trims. A cache read drops from $1.00 to $0.25 per million tokens, or 0.025 times the base input price. Anthropic estimates the effect at roughly 25% savings on a typical workload, and up to 45% on agentic usage, where re-read content dominates the bill. The same logic already ran through field reports on editorial agents in production: the cost centre is not generation, it is re-reading.
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-fable-5-1",
"max_tokens": 1024,
"system": [
{
"type": "text",
"text": "Stable context: specs, tools, codebase...",
"cache_control": { "type": "ephemeral" }
}
],
"messages": [
{ "role": "user", "content": "Question for the current turn" }
]
}'The block tagged cache_control is billed at the write rate on the first call, then as a read (now $0.25 per million) on each following turn as long as it stays unchanged. The longer the loop and the heavier the context, the more the cut matters.
What the raw numbers say
The base sheet does not move: input stays at $10 and output at $50 per million tokens. Only the cache read is revised.
| Line item | Price per million tokens |
|---|---|
| Input | $10 |
| Output | $50 |
| Cache read (Fable 5) | $1.00 |
| Cache read (Fable 5.1) | $0.25 |
On capability, Anthropic highlights a 52.6% score on Terminal-Bench-Science 0.1, against 24.7% for Fable 5 and 29.0% for Opus 5. The claimed gain is clear, but it rests on an in-house benchmark aimed at solving scientific problems in a terminal.
| Model | Terminal-Bench-Science 0.1 |
|---|---|
| Fable 5 | 24.7% |
| Opus 5 | 29.0% |
| Fable 5.1 | 52.6% |
On agentic workloads, a cache read at $0.25 per million tokens weighs more on the bill than a benchmark point.
Worth watching. The 52.6% figure comes from a benchmark published by the vendor, and always-on adaptive thinking can inflate the number of output tokens, which are billed at the full rate. The announced saving depends on how much is actually cached: a measurement on the production workload remains essential before rolling a model out to an entire fleet of agents.
The takeaway
Fable 5.1 does not change the input or output price; it makes re-reading context four times cheaper. For a brochure site that calls a model now and then, the effect is marginal. For a team operating long-context agents with many loops, the cut moves the break-even point and makes architectures that were too expensive suddenly viable.
On my own agent pipelines, cache reads already accounted for more than half of the monthly bill, so I am watching this cut with very concrete interest. Still, I am wary of the “up to 45%” savings headline: it assumes a caching discipline few teams actually maintain. Before migrating an entire fleet, I measure the real cache-hit rate over a week of traffic; that is what decides, not the spec sheet. — Simon Janvier
Further reading
Model details and pricing on the primary source: Anthropic newsroom.
