AI SpaceSep 1, 20268 min read

Claude Fable 5.1: The Cache Price Moves More Money Than the Benchmarks

Fable 5.1 doubles its predecessor on two hard benchmarks. The change that shows up on your invoice is a 75% cut to cache reads.

On this page
  1. What moved
  2. The pricing change, and why it lands harder
  3. What it costs to finish a task
  4. Where it sits against the open field
  5. How we would use it
  6. Sources

Anthropic shipped Claude Fable 5.1 on 1 September 2026, alongside Claude Mythos 5.1. Same underlying model, different safety posture: Fable 5.1 is generally available as claude-fable-5-1, while Mythos 5.1 goes to vetted US organisations through Project Glasswing.

The benchmark jumps are large and worth walking through. The line that will change more budgets is four words in the pricing section: cache reads cost 75% less.

What moved

Fable 5 to Fable 5.1 — % of tasks completed
Fable 5 to Fable 5.1 — % of tasks completed
ModelTerminal-Bench 4.0Terminal-Bench-Science 0.1
Claude Fable 542%24.7%
Claude Fable 5.155.8%52.6%

Sources: MarkTechPost: Claude Fable 5.1 and Mythos 5.1

Terminal-Bench-Science 0.1 went from 24.7% to 52.6%, and Terminal-Bench 4.0 from 42.0% to 55.8%. Doubling on a benchmark built to be hard is a genuine result, and Terminal-Bench-Science is one of the harder ones, Opus 5 sits at 29.0% on it.

The rest of the card, for completeness:

BenchmarkFable 5.1
Terminal-Bench-Science 0.152.6%
Terminal-Bench 4.055.8%
CursorBench 3.2.073.4%
Humanity's Last Exam, no tools60.9%
Humanity's Last Exam, with tools65.0%
AutomationBench31.4%
OSWorld 2.0, strict41.7%
GDPval-AA v21853

A note on reading these, because it applies to every model this season: Terminal-Bench 4.0 and Terminal-Bench-Science 0.1 are separate instruments from the 2.1 that most open models still report. Lining a 55.8 up against an 88.2 from another card produces a number and no meaning. We wrote about why the version number is part of the score.

The pricing change, and why it lands harder

Base pricing is unchanged at $10 per million input tokens and $50 per million output. What changed is the cache read: $1.00 down to $0.25.

Anthropic puts the effect at roughly 25% off typical workloads and up to 45% off highly agentic ones. That second figure is the interesting one, and the reason for the spread is structural.

Where cache reads land in an agentic runThe first turn writes the repository context to cache at full price. Every turn after that reads the same prefix back at the cache rate, so on a long run the cache read is most of the input bill.Turn 1: write 250Kof repo contextFull prefix + the turn's questionDiffTurn 2: reuse the same prefixCached prefix +the turn's questionTurns 2..N read the prefix at the cache rateDiffHarnessPrompt cacheModelLEGENDCallResponse

An agentic coding run re-sends the same repository context on every turn. Turn one writes that context to cache. Turns two through forty read it back. On a long run, cache reads are not a rounding error on the bill, they are most of the input line.

Cutting that line by three quarters is worth more to an agent that runs for an hour than any benchmark point, because it applies to every turn.

What it costs to finish a task

Take the same profile we use across these comparisons — 250K input tokens of accumulated repository context and 30K output tokens — and assume a well-cached agentic run where 90% of input comes from cache:

Fable 5Fable 5.1
Input, 225K cached$0.225$0.056
Input, 25K fresh$0.250$0.250
Output, 30K$1.500$1.500
One attempt$1.98$1.81

On a single attempt the saving looks modest, because output dominates. The saving grows with the number of turns that reuse the cache, which is exactly the workload Anthropic is naming: long, tool-heavy, many-turn runs where the same context is read back dozens of times.

So the practical guidance is narrow and useful. If your agent makes a handful of calls with fresh context each time, this release is a benchmark upgrade. If it runs long sessions over a stable repository, it is a real price cut, and the more agentic your setup the more of that 45% you will see.

Where it sits against the open field

Direct comparison is limited by the version problem above: the open models report Terminal-Bench 2.1 and 3.0, while Fable 5.1 reports 4.0 and Science 0.1. On the metrics that do line up:

MetricClaude Fable 5.1GLM-5.3GLM-5.3-Flash
Context window1M tokens1M tokens1M tokens
Native visionYesNoYes
LicenseProprietaryOpen weights (GLM-5.3 licence)Open weights (MIT)
Input $/1M$10$1.40$0.15
Output $/1M$50$4.40$0.50
HLE (with tools)65%62.5%55.3%
AutomationBench31.4%48.2%48.8%
GDPval-AA v218531769

† vendor-reported (self-reported by the model's vendor, not an independent harness). vendor labels 'HLE'; tools qualifier not stated. AutomationBench v1.0.6.

Sources: MarkTechPost: Claude Fable 5.1 and Mythos 5.1, GLM-5.3 model card (Hugging Face), MarkTechPost: GLM-5.3-Flash release, Z.ai GLM-5.3-Flash docs

Two things stand out. On Humanity's Last Exam with tools, Fable 5.1's 65.0% leads GLM-5.3's 62.5%, a real but narrow margin on a genuinely hard reasoning test. On AutomationBench, GLM-5.3 reports 48.2% and GLM-5.3-Flash 48.8% against Fable 5.1's 31.4%, with the caveat that GLM's figures come from AutomationBench v1.0.6 and the harnesses differ.

The price columns are the ones to sit with. Fable 5.1 lists at $10 and $50; GLM-5.3-Flash at $0.15 and $0.50, with open MIT weights and a 1M context.

How we would use it

Fable 5.1 has earned a specific job rather than a general one. It is the model for work where the difficulty is real and the cost of a wrong answer is high: the hard debugging session, the migration nobody has done before, the reasoning task that a cheaper model has already failed twice.

Put it at the end of a routing chain rather than the front. Let a cheap open model take the first pass at the routine majority, keep a verifier between the model and the merge, and escalate to Fable 5.1 for the cases that survive. That way you pay $10 per million input only on work that has demonstrated it needs the money, and the cache discount applies to the long sessions where it compounds.

For the arithmetic behind that pattern, see cost per solved task. For the setup itself, see running open models at frontier level.

Sources

Written by

Cho Yin Yong

Principal AI Solutions Engineer, XY Space

Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.

More from Cho Yin Yong

Share this article

Work with us

We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.

Start a project
Work with us

Book a call.We'll come back with specifics.

Start with the map of your organization, or with the one job that hurts. Measured in hours and money, and everything we build stays yours.

Loading form…