Scale · founder · 8 min read

Claude Fable 5.1: The Price Cut Hiding Behind an Unchanged Price Tag

Anthropic kept Fable's sticker price identical and cut the real cost 25-45%. Why your tool bill may move before your tool announces anything.

Published September 11, 2026 ·

AI models Claude Fable 5.1 pricing agents costs

Anthropic shipped Claude Fable 5.1 on September 1, three months after Fable 5. The announcement reads like a routine point release. The sticker price is identical — $10 per million input tokens, $50 per million output. The version number has a .1 in it.

Ignore all of that. The number that matters is buried two sections down in the pricing table, and it moved 75%.

The actual change: cache reads went from $1.00 to $0.25

When an AI agent works on your project, it re-reads the same things constantly — your codebase, your instructions, the tool definitions, the conversation so far. Model providers cache that repeated context so you’re not charged full price to send it again. On most Claude models a cache read costs 10% of the normal input price.

Fable 5.1 charges 2.5%. A cached read is now $0.25 per million tokens, down from $1.00 on Fable 5.

That produces a genuinely odd price profile. Fable 5.1’s fresh input and output cost twice what Opus 5 costs. But its cached input costs half what Opus 5’s does, and it’s only 25% above Sonnet 5 — a model whose base input price is five times cheaper.

Anthropic’s own figure is that this makes Fable 5.1 roughly 25% cheaper for typical workloads and up to 45% cheaper for heavily agentic ones, measured across four weeks of August usage. The more your work looks like “an agent grinding on one codebase for an hour,” the bigger the discount.

Why this is a founder problem, not a developer problem

You probably don’t buy tokens directly. You buy Lovable, Cursor, Replit, or Claude Code, and someone else buys the tokens.

That’s exactly why it matters. When the economics of a model change this sharply, the tools sitting on top of it re-route their traffic — and they don’t usually send you an email about it. Cognition said the quiet part out loud on launch day: Walden Yan confirmed Devin was moving its Opus 5 traffic to Fable 5.1 immediately, starting with code review, because “a Fable-class model is finally economical for the workloads we’d kept on Opus.”

So the practical read is this. If a tool you pay for feels different this month — better on long tasks, or noticeably faster to burn through your credits — a model swap you were never told about is a plausible explanation. That’s not a scandal. It’s just how this layer of the stack works now, and it’s worth knowing so you don’t misdiagnose it as your own prompting getting worse.

The benchmarks, with the caveats attached

Anthropic published seven benchmark rows. Fable 5.1 leads Opus 5 on all seven. The two worth your attention:

• Terminal-Bench-Science 0.1 — 52.6%, against 24.7% for Fable 5 and 29.0% for Opus 5. More than double, inside a point release. The error bars are ±3.5 to 4.5 points, so the doubling survives them.
• CursorBench 3.2.0, the closest thing here to “can it actually code” — 73.4%, against 70.5% for Fable 5 and 70.0% for Opus 5.

Notice the gap between those two rows. The science benchmark more than doubled. The coding benchmark moved three points. Coding is the row where Fable 5.1’s lead is thinnest, which tells you something about where the frontier is now: the labs have largely stopped competing on whether the model can write a function, and started competing on whether it can run a four-hour investigation without a human.

Two caveats you should hold onto:

These are vendor-reported numbers from Anthropic’s own evaluation setup. Anthropic is unusually transparent about the qualifications — it flags that its production safeguards score a flat zero wherever they intervene, which likely understates its own models on the computer-use and business-workflow rows. But they’re still the vendor’s numbers.

And on business workflows, the honest headline isn’t the near-doubling. It’s that the score is 31.4%. Autonomous business workflows remain unsolved, by everyone, including the model that leads the table.

About the Astra comparison

This site has said in a few places that Fable 5.1 is ahead of GPT-6 Astra. That claim needs a footnote.

Anthropic’s comparison table doesn’t contain Astra. It compares against GPT-5.6 Sol, which Fable 5.1 leads on all five rows where Sol appears — and Sol doesn’t appear at all on computer use or reasoning. Astra launched on September 3, two days after this table was published.

So the honest version is: Fable 5.1 beats OpenAI’s previous flagship on every published head-to-head, and leads Astra on the two independent-ish scores OpenAI itself released. Nobody has run them against each other on the same harness. If you see a confident ranking of these two models anywhere, including here, treat it as an inference rather than a measurement.

What to actually do about it

If you’re building with a hosted tool, do nothing. You don’t choose the model; your tool does. Just know that the price change makes long-running agent work cheaper for your vendor, which over the next quarter usually shows up as more generous limits or more ambitious agent features.

If you pay for API usage directly, check your effort setting. Anthropic claims Low and Medium effort on 5.1 now match Fable 5’s full-effort results at meaningfully lower cost. Defaults are High in Claude Code and Medium in Claude Cowork and on claude.ai. If you’re on High out of habit rather than need, that’s the cheapest thing you can change today.

Don’t switch tools over this. A three-point move on a coding benchmark is not a reason to migrate a working project. The thing that actually determines your outcome is still whether you can describe what you want clearly and tell when the result is wrong.

The bottom line

Fable 5.1 is the first frontier release in a while where the interesting number is a price, not a score. Anthropic had a demand problem — a Financial Times report found Fable 5 accounted for only about 11% of Anthropic model spending two months after launch, with cheaper models taking share — and it solved that by making the expensive model cheap in exactly the scenario where it’s used most.

For you, the takeaway is smaller and more useful than the launch coverage suggests: cost per finished task is now the number that matters, and it is no longer predictable from the price per token. A model that’s twice as expensive on paper can be cheaper in practice for agent work, and your tools are already making that calculation on your behalf.

Sources: VentureBeat, Vellum, Anthropic.