Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding

Fable 5.2 vs Opus 5.5 on real coding work

Written November 10, 2026

Anthropic says Fable 5.2 matches Opus 5.5 at about half the cost per task. One independent test on six real C# tasks found the opposite on price: cleaner code, zero bugs, and two to three times the cost.

The problem

A coder choosing between the new Fable and the known Opus 5.5 has two claims in front of them that point in opposite directions. The vendor's claim comes from its own coding benchmark. The independent claim comes from six real tasks, run dozens of times each, with every result built, tested, and load-tested by the author. Both are worth carrying, and neither settles the question on its own.

The test

Rajni Kavathiya's test, published November 10, took six .NET 10 tasks she had already run 24 times on Opus 5.5 in Claude Code and ran them 36 more times on Fable 5.2: 24 runs at High effort and 12 at Medium, against the 24 Opus 5.5 runs at Medium effort. She built, tested, and load-tested every result herself. The effort levels were not matched, which matters for the cost comparison below.

What she found

MeasureFable 5.2Opus 5.5
Runs3624
Bugs found on review05
Cost per task2–3x Opusbaseline
One endpoint under load (200 requests)194 or more answered142 failed

The load result is the sharpest difference: every Opus version of one endpoint failed 142 of 200 requests under load, while every Fable version answered 194 or more. Fable also found a new way to leave a server running, which the author counts against it. And on five of the six tasks, she writes, the extra money bought nothing she could measure.

The cost conflict

Anthropic's announcement says that on its FrontierCode 2.0 evaluation, Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, for roughly half the cost per task. The independent test says the opposite on its own tasks: two to three times the cost per task. The two claims measure different things on different work, Fable ran at higher effort in the independent test, and the per-task price of a model depends on how many tokens the task takes, not only on the price per million tokens. There is no way to reconcile them from the published numbers, so this page carries both and picks neither.

Verdict

On six real C# tasks, Fable wrote cleaner code at two to three times the cost. That is the author's verdict, from her test, not a general finding about the two models.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Claude Fable 5.2 / Mythos 5.2: what changed · Claude Code 2.1.321: Fable 5.2 is default