TokenSave Data · Dataset

Claude Fable 5.2 costs 40% less than 5.1 per run, not 20%, and fixed the same bugs as Opus 5.5 at 83% more: 125 runs

Claude Fable 5.2 cost 40% less per run than Fable 5.1 ($0.73 vs $1.22), twice its 20% list-price cut. It also fixed exactly as many of 25 real bugs cleanly as Opus 5.5: 37 of 50 runs each. But Opus 5.5 did it at $0.40 a run, so Fable 5.2 cost 83% more for the same fixes.

Measured 2026-11-13 · 125 runs in the plan (135 published) · 25 real bugs · License CC BY 4.0

DOI 10.5281/zenodo.24186078 · Also on Hugging Face · Kaggle · GitHub

Key findings

• Fable 5.2 vs 5.1, cost: -39.9% per run (95% bug-level interval -49.8% to -27.2%), against a list-price change of -20%. Fable 5.2 wrote 27% fewer output tokens, read 67% less cache, took 14.6 turns instead of 21.0, and finished in a median 65 s instead of 136 s.
• Fable 5.2 vs Opus 5.5, fixes: +0 points (95% interval -8 to +8; pre-registered threshold 15 points). Both had 37 clean fixes in 50 runs: 26/30 on the easy bugs and 11/20 on the hard ones.
• Cost per clean fix: Opus 5.5 $0.54, Fable 5.2 $0.99, Fable 5.1 $1.60.
• Fable 5.1 fixed 19 of 25 cleanly (1 run per bug), close to Fable 5.2.
• All three missed the same four bugs (cachetools class access, js-yaml BOM, tomlkit underscores, zod numeric keys). None found what the report left out.

Results (125 runs in the plan)

Clean fix = every maintainer test for the fix passes and no other test breaks. Cost is Claude Code’s own per-run figure at API list prices.

Easy bugs: Fable 5.2 $0.48, Fable 5.1 $0.67, Opus 5.5 $0.26 per run. Hard bugs: $1.11, $2.04, $0.61.

Bug by bug

Method

• Pre-registered on 2026-11-13: PREREG.md, with one dated amendment.
• Bugs and scoring from dataset 25: 25 real bugs fixed upstream in August to October 2026 in 13 open-source projects (15 easy, 10 hard), each run starting at the commit before the fix, scored with the maintainers’ tests by the unchanged dataset 25 scorers.
• Agent: Claude Code 2.1.322, headless claude -p, effort medium, acceptEdits, tests and builds allowed, no installs, no network, no web tools, Windows 11.
• Amendment 1: the first 21 runs used the subscription’s whole 5-hour window in 34 minutes, so before any scoring the plan was cut from 175 to 125 runs (Fable 5.2 × 2, Fable 5.1 × 1, Opus 5.5 × 2 per bug). Ten runs already finished in the dropped cells are published but left out of the analysis.
• Incidents: 6 runs cut off by the 5-hour limit and 2 by a desktop-app restart were re-run; the cut-off attempts are not scored.

Limitations

• 25 bugs, 2 runs per bug for Fable 5.2 and Opus 5.5 and 1 for Fable 5.1, so the Fable 5.1 fix rate is the least certain number.
• One machine, one prompt style, effort medium only. Bounded single-issue bugs; larger multi-file work may separate the models.

How to cite

Free to use, quote and republish with attribution under CC BY 4.0. Please link to this page.

BibTeX

@misc{tokensave_claude_fable_5_2_real_bugs_2026,
  author    = {{TokenSave}},
  title     = {Claude Fable 5.2 costs 40\% less than 5.1 per run, not 20\%, and fixed the same bugs as Opus 5.5 at 83\% more: 125 runs},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.24186078},
  url       = {https://doi.org/10.5281/zenodo.24186078},
  note      = {Dataset, CC BY 4.0}
}

Embed the chart

Related

• Claude Haiku vs Opus on 25 real bugs (dataset 25)
• Explore subagent model by Claude Code version (dataset 28)
• Estimate your own Claude Code bill

Questions or corrections: contact@jonhisking.com