# Why Claude Fable 5.2 cost four times Opus 5.5 for the same score: a token-by-token breakdown of The New Stack's tests

> Source: <https://theainewsreport.com/2026-11-16-claude-fable-5-2-vs-opus-5-5-cost-breakdown.html>
> Published: 2026-11-16 15:43:45+00:00

# Why Claude Fable 5.2 cost four times Opus 5.5 for the same score: a token-by-token breakdown of The New Stack's tests

The New Stack ran Claude Fable 5.2 and Opus 5.5 on nine coding and agent jobs and got the same perfect scores at very different prices. We split the published token counts to show the gap is mostly reasoning tokens billed at twice the price, and what a fair rerun would change.

**This explains reporting by**

[The New Stack, Jessica Wachtel: Claude Fable 5.2 vs. Opus 5.5: Same scores at almost 4x the cost (November 15, 2026)](https://thenewstack.io/claude-fable-5-2-vs-opus-5-5/).
Read the original first:

[https://thenewstack.io/claude-fable-5-2-vs-opus-5-5/](https://thenewstack.io/claude-fable-5-2-vs-opus-5-5/)

## In one minute

- The New Stack gave both models three jobs in three versions each. Both scored 150 of 150 in every group.
- Fable 5.2 cost an estimated $8.98 and took 1,579 seconds; Opus 5.5 cost $2.29 and took 701 seconds.
- By our count of the published numbers, most of the gap is reasoning: Fable wrote 3.4 times as many reasoning tokens, and each output token costs twice as much ($40 against $20 per million).
- The models ran at their API defaults, which differ: Fable at high effort, Opus at medium. The author says so; it is the first thing a rerun should fix.
- Anthropic itself recommends starting most coding and knowledge work on Opus 5.5.

## What The New Stack tested

The author ran each model once per test through Anthropic's API, at default settings, with no turn cap, and graded answers with a script. The three jobs: write a cron scheduler from a spec and absorb a change request (50 hidden tests), fix interacting bugs in a small invoicing library while reporting one wrong test without editing it, and find the cause of a payments outage in logs that include decoy deploys and servers with wrong clocks.

Both models scored perfectly on the first version of each job, so the author made each one harder twice. Both still scored 50 of 50 on all nine runs.

For context, Anthropic's own benchmarks, as the article reports them, favor Fable: 49.1% against 45.5% on FrontierCode 2.0 and 69.9% against 62.3% on FrontierSWE v2. The article's tests are much smaller and, as the author says, cannot prove or disprove those scores.

## Where the money went

We added up the per-run figures the article publishes. Totals match the article's table.

- Fable 5.2: 722,188 tokens in, 168,099 out, of which 91,615 were reasoning. Estimated cost $8.98.
- Opus 5.5: 637,949 tokens in, 78,293 out, of which 26,908 were reasoning. Estimated cost $2.29.
- Reasoning tokens: Fable wrote 3.4 times as many. Output that was not reasoning: only 1.5 times as many (76,484 against 51,385).
- Output at list price: 168,099 tokens at $40 per million is $6.72, about 75% of Fable's total. 78,293 at $20 per million is $1.57, about 68% of Opus's.
- Input: at the uncached list price, input alone would be $5.78 for Fable and $2.55 for Opus, which would push the totals to $12.50 and $4.12. The article's lower totals mean much of the input was billed as cache reads, which Anthropic prices separately.

So the four-to-one gap is mostly two things multiplied: more thinking, and a higher price for each token of thinking. Input volume was close between the two.

## The gap is not the same on every job

- Cron scheduler: Fable $3.29, Opus $0.99. Fable wrote two to three times the reasoning on every version.
- Invoicing repo: Fable $3.15, Opus $0.39, the widest gap. Opus finished each version in under a minute.
- Incident investigation: Fable $2.54, Opus $0.91. Here Fable took fewer turns (9 to 11 against 12 to 14) and still cost more.

The closest result came on the job that most rewards careful digging. That fits Anthropic's own advice that Fable is for more open-ended technical work.

## The effort confound

Anthropic's API defaults differ: Fable 5.2 runs at high effort and Opus 5.5 at medium. The effort setting controls how much the model thinks before it answers, so it acts directly on the reasoning tokens that drive this gap.

That makes the test a fair picture of what a developer gets without tuning anything, which is how many teams start. It does not tell you how the two compare at the same setting. Anthropic's guidance, as The Decoder reports it, is that Opus 5.5 on Medium matched Opus 5 on High in its internal tests, and that Medium is the place to start.

A fair rerun would set both models to the same effort level, run each job several times rather than once, and report cache reads and uncached input separately.

## How to run your own version

- Pick three jobs from your own backlog: one fix, one review, one build from a spec.
- Run each model at its default, then both at the same effort setting.
- Log input, cached input, output and reasoning tokens for every run, not only the bill.
- Compare cost per passed task. Move up to the pricier model only on the job types where the cheaper one fails.

## Who is affected

| Case | Status | 
|---|---|
| Teams paying per token for coding agents | On these tests, the same score cost about four times as much on Fable 5.2. | 
| Teams already on Opus 5.5 at medium effort | The test supports staying there for well-scoped coding work. | 
| Open-ended technical work | Not covered by these tests. Anthropic recommends Fable for it. | 

## What to do

- Start coding and agent work on Opus 5.5 at medium effort, and track cost per passed task.
- Try Fable 5.2 only on the job types where Opus falls short, and compare at the same effort setting.
- Log reasoning tokens separately, since they drove most of this gap.

## What is still unknown

- Each test ran once per model, so run-to-run variation is unknown.
- The split between cache reads, cache writes and uncached input was not published, so we cannot rebuild the exact dollar figures, only bound them.
- How the two compare at the same effort setting. The article tested defaults only.
- Anthropic's benchmark scores quoted in the article are Anthropic's own.

## Sources

- [The New Stack, Jessica Wachtel: Claude Fable 5.2 vs. Opus 5.5: Same scores at almost 4x the cost (November 15, 2026)](https://thenewstack.io/claude-fable-5-2-vs-opus-5-5/) — the original report
- [Anthropic: Claude Fable 5.2 and Claude Mythos 5.2 (November 9, 2026)](https://www.anthropic.com/claude-fable-and-mythos-5-2)
- [The Decoder: Anthropic says outdated prompting habits are slowing down Claude Opus 5.5 (November 14, 2026)](https://the-decoder.com/anthropic-says-outdated-prompting-habits-are-slowing-down-claude-opus-5-5/)

[AI News Report](https://theainewsreport.com/)· every headline, every morning.
