Anthropic’s Fable 5.1 Discount Lives Only in Cache Reads

Anthropic left Claude Fable 5.1’s list price at $10 and $50 per million tokens, then called the model cheaper. The company said it should deliver estimated 25 percent typical savings, and up to about 45 percent on highly agent work, wherever usage is billed by token.

The only rate that moved is cache reads, now $0.25 per million tokens. Two labs already posted opposite bills for the same launch.

The $10 and $50 Rates Did Not Move

Claude Fable 5.1 reached the API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude’s AWS platform on September 1, 2026. Input stays $10 per million tokens and output stays $50, the same sticker Fable 5 carried, and twice Claude Opus 5’s $5 and $25.

Anthropic cut cache reads by 75 percent, from $1.00 to $0.25. Cache writes stay $12.50 per million tokens for the five-minute tier and $20 for the one-hour tier. Batch jobs still get 50 percent off input and output. The Fable 5.1 API price sheet lists a 1 million token context window, 128K max output, and a June 2026 knowledge cutoff.

FABLE 5.1 TOKEN RATES

That cache line is now 2.5 percent of Fable’s input rate. Most frontier APIs still bill cached input at about 10 percent of the raw input price, which is why Fable’s rereads undercut Opus 5’s $0.50 cache even though Fable’s list price is double.

Anthropic’s 25 and 45 percent figures come from four weeks of its own August 2026 usage at default effort, covering Claude Enterprise, Claude Code, and the API. On that mix, a typical workload falls from an index of 100 to about 75. A highly agent workload, where cache reads are most of the cost, falls to about 55. A short prompt that never hits the cache pays the old $10 and $50 rates in full.

Cache reads with Fable 5.1 cost 75% less than Fable 5’s.

This reduces the cost of the model in practice by around 25% for typical workloads, and up to 45% for highly agentic ones. pic.twitter.com/uGjoKHcbOD

— Claude (@claudeai) September 1, 2026

Two Labs Measured Cost and Got Opposite Bills

Artificial Analysis, which helped Anthropic with pre-release tests, ran Fable 5.1 at max effort on its Intelligence Index. The model scored 66, ahead of Opus 5 at 63, Fable 5 at 62, and GPT-5.6 Sol at 61. The bill was $3.76 per index task, which the firm called 20 percent more per task than Fable 5’s $3.14, and about 1.6 times Opus 5’s $2.34.

The extra spend is output. Fable 5.1 used about 1.7 times as many output tokens as Fable 5 at max. The cheaper cache saved about $1.40 per task, mostly on agent-style evals where most input tokens are rereads. Without that cut the same run would have cost about $5.16. At extra-high effort the score is 65 at $2.72, which is $1.04 less than max and still above Opus 5.

Cognition, the company behind the Devin coding agent, published the other ledger. On FrontierCode 1.1 Extended at medium effort, Fable 5.1 cost $2.68 per task, down from Fable 5’s $5.84, a 54 percent cheaper per coding task result, with the score moving from 62.8 to 63.6. Opus 5 landed at $3.51. Devin users, Cognition said, should see 10 to 25 percent savings on real work.

COST PER COMPLETED TASK

Both numbers can be right. Artificial Analysis is scoring a broad exam at max effort, where Fable 5.1 writes a lot. Cognition is scoring mergeable coding work at medium effort, where almost every token is a cache hit. Anthropic’s own 25 to 45 percent range sits in the middle, because it used default effort on mixed production traffic, not a single lab harness.

Why Cache Reads Eat an Agent’s Invoice

An agent does not send a fresh prompt each turn. It rereads the repo, the task, the tool list, and its own earlier steps, then writes a short action. Cognition’s typical FrontierCode task on Fable 5.1 shows how lopsided that mix is, and why a cache cut moves the total when a sticker cut would not.

ONE FRONTIERCODE TASK ON FABLE 5.1

• Cached reads: About 3 million tokens, the pile that used to cost $3.08 at $1.00 and now costs $0.77 at $0.25.
• Fresh input: About 70 thousand uncached tokens, billed at the unchanged $10 rate, about $0.84 in Cognition’s breakdown.
• Output: About 21 thousand tokens at $50 per million, about $1.07.
• Share from cache: More than 95 percent of all tokens on both Fable 5.1 and Opus 5 are rereads of context the model has already seen.

At the old $1.00 cache rate that task cost about $4.99. At $0.25 it is $2.68. Opus 5 still rereads about 4.5 million cached tokens, writes about 26 thousand, and sends about 85 thousand fresh, and its $0.50 cache rate leaves it at $3.51. Fable 5.1 also used 33 percent fewer tokens than Opus 5 on the same suite, so the cheaper reread is not the only lever.

Cognition put the point without hedging. Token stickers, the firm said, are a poor way to talk about agent cost, and cost per finished task is the figure that matches how these systems actually run.

Coding Agents See the Real Price Cut

Walden Yan, Cognition’s co-founder and chief product officer, did not wait for a pricing debate. He said Devin would move Opus 5 traffic onto Fable 5.1 on launch day, because the new cache rate finally made a Fable-class model cheap enough for jobs the company had kept on Opus, starting with code review.

We’re moving our Opus 5 traffic in Devin to Claude Fable 5.1 on launch day. It matched or edged out Fable 5 in our testing at a lower cost per task, and with the new cache read pricing a Fable-class model is finally economical for the workloads we’d kept on Opus, starting with code review.

Walden Yan, Co-founder and CPO, Cognition

That is the customer Anthropic wrote the 45 percent line for: long loops, stable system prompts, tool schemas that never change, a repo that gets pulled in on every step. Claude Code is the obvious in-house version of that pattern, and Anthropic’s August sample included it.

Chat-shaped work is the other stack. A one-off question, a short rewrite, a single classification call, those jobs barely touch the cache, so they still pay $10 and $50. Builders who hit plan caps on Claude’s subscription tiers also do not pocket an API cache cut. The cheaper reread applies where usage is billed by token, which is the clause Anthropic put on the savings claim.

The quality pitch is easier to see on a single bug. Damien, a senior portfolio manager at Millennium, said Fable 5.1 found a crash that showed up about once in a million runs, after four to five years in which his team and every other model he tried, including Fable 5, missed it. The new model, he said, disassembled a vendor library, matched it to the core dump, and traced the fault there.

Max Effort Turns the Discount Into Extra Tokens

The trap sits in the effort knob. Anthropic said Fable 5.1 at Low or Medium effort matches or beats Fable 5 at a much lower cost. The API’s default effort is High. Claude Code also defaults to High. Claude Cowork and Claude.ai default to Medium. Anyone who points the new model at a hard job and leaves the default in place is closer to Artificial Analysis’s max-effort bill than to Cognition’s medium-effort coding bill.

EFFORT IS THE VOLUME DIAL

• Token span: Across five effort settings, Fable 5.1 used 13.1 million output tokens at low and 143.7 million at max on the Intelligence Index, an 11-times range.
• Score span: Those same settings scored from 58 to 66 on the index.
• Fallback share: About 4 percent of output tokens in the AA run were served by Opus 4.8 or Opus 5 after safety flags.
• Attempt rate: At max, it tried 93.4 percent of AA-Omniscience questions, against 87.8 percent for Opus 5, and it answered more of the ones it got wrong instead of declining.

More attempts can look like a better model and still print a larger invoice, because output is the $50 line. On Cognition’s own leaderboard the score peaks at medium and then falls below Fable 5 at higher effort, not because pass rate drops, but because the model starts editing files outside the task and fails a mergeability rule called scope.

Anthropic’s own benches still show a clear step up when the extra tokens buy the right work. Terminal-Bench-Science 0.1 goes from 24.7 percent on Fable 5 to 52.6 percent on Fable 5.1. Terminal-Bench 4.0 agent coding goes from 42.0 percent to 55.8 percent, with Mythos 5.1 at 60.9 percent. Those gains are why a shop might accept a higher max-effort bill. They are also why a shop that wanted the 45 percent headline has to pin effort, not just switch the model ID to claude-fable-5-1.

Fewer False Alarms, Same Mythos Gate

Fable 5.1 and Mythos 5.1 are the same weights with different safety filters. Fable is on every listed platform. Mythos stays inside trusted access programs, including Project Glasswing, for cyber defense and life-science work.

Anthropic said the new cyber filters flag benign requests about 60 percent less often, in part because Fable 5.1 may be used to find software bugs but not to build exploits. On basic biology and medical questions, the company said the fallback rate recently dropped by around 85 percent. Mythos keeps the tighter biology limits and will have a separate access program built with the U.S. government.

Jane Street’s head of quantitative research, Craig Falls, said the model solved more of the firm’s coding problems than Fable 5 or Opus 5, and that it stayed readable on long multi-step tasks where older models became hard to follow. That matches the product Anthropic is selling: a slow, expensive model that is meant to sit on a job for hours, reread its own notes, and keep going.

Enterprise Frontier Safeguards Arrive Later This Fall

The other customer complaint Anthropic listed next to price was data retention. Enterprise Frontier Safeguards, the company said, store data in cloud servers the customer controls, not at Anthropic, with privacy equivalent to a zero-data-retention policy while still watching for abuse. EFS will roll out in phases, starting later this fall. Until then, eligible customers can run Fable 5.1 with zero data retention, a mode Fable did not offer before.

Cognition said that exemption now covers Fable 5 and Fable 5.1 for eligible Devin customers, and that it is feeding Anthropic notes on the EFS design so the coding agent can keep frontier models without parking customer code on Anthropic’s disks. For a bank or a lab that would not send repos to Fable at all, that permission is the thing that makes the $0.25 cache rate usable.

The list price is still $10 and $50. The cheaper reread is real for loops that reuse context, and the extra output is real when effort is left on high. Teams that pin medium effort on cached coding work are already posting lower task bills. Teams that treat the 45 percent line as a sticker change will see the old invoice, plus a more talkative model.