Claude Fable 5.1:
the complete guide
Released 2026-09-01. Input at $10 and output at $50 per million tokens are unchanged from Fable 5; what moved is cache reads, from $1.00 to $0.25 — the announcement says "Cache reads now cost 75% less". Context is 1M and max output 128K. On QCode the model id is claude-fable-5-1.
Four pricing and spec figures
Cache read, per million tokens
Fable 5 charges $1.00, so this is 75% lower. Workloads that re-read long context repeatedly benefit most directly.
Input / output, per million tokens
Identical to Fable 5 — the list price did not move. Only the cache line changed.
Context window / max output
Same as Fable 5. Cache writes also unchanged at $12.50 for the 5-minute tier and $20 for the 1-hour tier.
Official estimate for a typical bill
The announcement puts typical workloads around 25% cheaper than Fable 5, with larger savings on complex, highly agentic tasks.
What actually changed
The headline rates did not move: $10 input and $50 output per million tokens, exactly as Fable 5. What moved is the cache read, from $1.00 to $0.25. That is not a marketing number — it decides the real bill for the "long context, many turns" class of workload. An agent re-reads its whole context on every turn, so cache reads are often the largest line on the invoice; cutting three quarters off that line does far more than trimming the input rate would.
Official wording and rates (verified 2026-09-02)
The announcement states "Cache reads now cost 75% less, or $0.25 per million tokens" and that "costs are reduced by around 25% relative to Fable 5", with larger savings on complex, highly agentic tasks. The rate table agrees line by line: input $10, output $50, cache read $0.25, cache writes $12.50 for the 5-minute tier and $20 for the 1-hour tier, 1M context and 128K max output. Mythos 5.1, released alongside it, shares the same base with looser guardrails and is available only to vetted organisations — it is not a publicly callable model.
Timeline of the Fable line
Fable 5 and Mythos 5 released.
Fable 5 was disabled for all customers under a US export control directive and restored on 1 July. Other Claude models were unaffected at the time.
Fable 5.1 and Mythos 5.1 released. Fable 5.1 is broadly available; Mythos 5.1 is restricted to vetted organisations.
Confirmed vs what you have to measure
Confirmed, with a first-hand source
① Cache reads moved from $1.00 to $0.25, stated as 75% lower in the announcement. ② Input $10 and output $50 unchanged. ③ 1M context and 128K max output. ④ The official estimate for typical workloads is around 25% cheaper. ⑤ Mythos 5.1 is not a publicly callable model.
What you have to measure
How much you actually save depends on your cache hit rate, which depends on your prompt structure and call cadence. The official 25% is an estimate for typical workloads, not a commitment. To measure: send one fixed, reasonably long input at least three times, drop the first, and look at what share of usage on the rest is cache reads — that share is what decides how much this price cut is worth to you.
Fable 5.1 vs Fable 5
A different bill shape
Input and output cost the same on both; the whole difference sits in cache reads ($0.25 versus $1.00). The higher your hit rate, the longer your context and the more turns you run, the larger 5.1's advantage. A one-shot short question shows almost no difference.
Specs and migration
1M context and 128K max output are identical, so switching is a matter of changing the model id. Note that 5.1 has some developer-side behaviour changes — for example how thinking content from earlier turns is handled — so run your own regression cases before migrating a long-chain agent.
How to decide whether the switch is worth it
① Measure what share of your current bill is cache reads — this needs no model change, just a look at the composition of the bill you already have. ② If that share is small (short conversations, a different input every time), this price cut barely touches you and there is no urgency. ③ If it is large (long-context agents, repeatedly reading the same codebase), your saving is essentially that line multiplied by 75%. ④ Run your own regression cases before switching; do not decide on price alone.
On QCode
The model id is claude-fable-5-1 (the dotted form claude-fable-5.1 normalises to the same model). Pricing is vendor list rate times a service multiplier, so the cache read reduction flows straight into your deduction — the deduction basis is the usage block in the response. To verify: send a fixed input a few times and check whether the cache read line in usage is being charged at the $0.25 tier.
FAQ
How much cheaper is Fable 5.1 than Fable 5?
On list price, only the cache read changed: $1.00 to $0.25 per million tokens, a 75% cut. Input at $10 and output at $50 are unchanged. The official overall estimate is around 25% cheaper for typical workloads and more for complex agentic ones — but the real figure depends entirely on your cache hit rate.
Did the context window or max output change?
No. The context window is 1M and max output 128K, the same as Fable 5. The two cache write tiers ($12.50 for 5 minutes, $20 for an hour) are also unchanged.
Can I call Mythos 5.1?
No. Mythos 5.1 shares a base with Fable 5.1 but has looser guardrails and is open only to vetted organisations; it is not a publicly callable model. Any channel claiming it can provision Mythos for you is worth asking for evidence.
Which model id should I write?
Write claude-fable-5-1. The dotted form normalises to the same model, but standardising on the hyphenated form keeps your own logs and bills from carrying two spellings of one thing.
Do I need code changes to migrate from Fable 5?
At the interface level, changing the model id is enough. But 5.1 carries some developer-side behaviour changes — for instance how thinking content from earlier turns is treated — so run your own regression cases before switching a long-chain agent wholesale on price alone.
How do I confirm the cut actually reached my bill?
Look at usage. Send one reasonably long fixed input at least three times, drop the first (a cold start), then take the cache read token counts and the matching deductions on the rest and work backwards to the unit price to see whether it lands on the $0.25 tier. That is more direct than any second-hand account.
Sources
The cache read cut and the overall reduction: Anthropic's official announcement page (fetched 2026-09-02; it contains "Cache reads now cost 75% less, or $0.25 per million tokens" and "costs are reduced by around 25% relative to Fable 5"). Per-line rates, context and max output: the litellm rate snapshot that our pricing is based on (checked 2026-09-02: input 1e-05, output 5e-05, cache read 2.5e-07, cache writes 1.25e-05 and 2e-05 per token; max_input 1,000,000, max_output 128,000). Mythos 5.1's availability is stated in the same announcement. Prices move with upstream changes; the official pages and our pricing page govern.
Measure your cache share first
The larger the cache read share of your bill, the more this cut is worth; if it is small, there is no rush to switch.
Further reading
Claude Fable 5 complete guide
The previous generation's specs, pricing and where it fits.
Claude Fable 5 pricing guide
Per-line rates and how cache pricing is calculated.
How to judge whether an AI API relay is worth using
Includes the cache hit rate measurement — the same test that tells you what this price cut is worth.
Prices and specs on this page were verified on 2026-09-02 and move with upstream changes. The official cost reduction is an estimate for typical workloads, not a commitment for yours. Model availability is governed by /models and your own testing.