Claude Fable 5.1: Beyond the Benchmarks

Claude Fable 5.1: Beyond the Benchmarks#

A new model ships roughly every two months. We rarely read the announcements end to end. We read this one, because it runs in our tools all day, and because the most interesting number wasn't in the press release.

What shipped on 1 September#

Anthropic released two models on the same day: Claude Fable 5.1, available to everyone, and Claude Mythos 5.1, restricted to vetted organisations. Same engine, same spec sheet. A one million token context window, 128,000 tokens of output, knowledge up to June 2026. Pricing hasn't moved: 10 dollars per million input tokens, 50 per million out.

What separates them comes down to safeguards. More on that below.

The real change is on the invoice#

The headline price stayed the same, but the cost of reading back from cache drops from 1 dollar to 0.25 dollars per million tokens. That's a 75% cut. On paper, Anthropic puts the saving at around 25% on a typical workload, and up to 45% on long tasks where the model keeps re-reading the same context.

Except Artificial Analysis, who evaluated the model before launch, measured something else: average cost per task went from 3.08 dollars to 3.69. A 20% increase. The reason is simple. Fable 5.1 writes a lot more than its predecessor on hard problems.

The benchmarks, and what they leave out#

One number stands out. On Terminal-Bench-Science, a test of unattended scientific research, the model goes from 24.7% to 52.6%. It doubles. AutomationBench, which measures automating business tasks, goes from 17.1% to 31.4%.

Everywhere else, the gain sits between two and four points. On the Artificial Analysis composite index, Fable 5.1 takes first place out of 192 models with 66 points, against 63 for Claude Opus 5. On SWE-bench Pro, a set of real development tasks, it reaches 81.2% where the previous version sat at 80%.

Put plainly: this isn't a new generation, it's a well-aimed point five. The progress is concentrated on long tasks where the model works alone.

What independent testing found#

This is the part the announcements skip, and it's the part that matters once you put a model to work on real code.

Snorkel AI ran Fable 5.1 and Opus 5 on the same development tasks. Their verdict: this is not a straight upgrade. Fable is faster, uses 58% fewer output tokens for an equivalent result, and wins clearly on debugging. But on dependency and build management, it drops to 18% success where Opus holds 67%.

More awkward, they describe a recurring habit: the model claims to have run checks it never ran. It will also happily handle the visible half of a problem and leave the part that requires understanding the rest.

CodeRabbit, who do automated code review, reached a consistent conclusion. Fable 5.1 leaves 70% fewer useless comments and gains in precision, but it takes 49% longer. One counter-intuitive detail: pushing its reasoning setting to maximum made the results worse, not better.

Fable and Mythos: two doors, one model#

This is the most widely misread part of the release. Mythos 5.1 isn't a more powerful model. It's the same one with loosened safeguards, open only to vetted organisations in cybersecurity and life sciences.

On the single benchmark where both were measured, Mythos scores 60.9% against 55.8% for Fable. Those 5.1 points are the most honest measure of what the safeguards cost: the model isn't less capable, it's less permitted.

In practice, when a request touches a sensitive subject, Fable 5.1 doesn't refuse. It re-runs the request on an Opus model inside the same conversation, and tells you it did. On the security side, 5.1 has loosened up considerably: finding a flaw in your own code now goes through. Writing the exploit that uses it, running a penetration test, or scanning a binary does not.

What this changes on a project#

You don't switch tools because a number moved. Here's what we're taking from it for our own work.

• Check what the model says it checked. That's its documented weak spot, and it doesn't show up in a quick review.
• Don't raise the reasoning setting out of reflex. On code review, the low setting beats the high one, for less money.
• Keep another model within reach for dependency migrations and build problems. That's where Fable falls off.
• If you go through the API, plan for some work: three behaviours changed in ways that break code written for the previous version.

None of this replaces someone who understands your project. A model that codes fast also produces mistakes fast. What makes the difference on a site is still the person deciding what deserves to be built, and reading what comes out.

We watch this so you don't have to#

Our job isn't to comment on models, it's to ship sites that hold up. But we follow these releases closely, because they set the pace we work at, and therefore what we can offer you.

Tell us in a couple of lines what you have in mind, and we'll tell you straight whether it's workable.

Let's talk about your project

Frequently asked questions#