Claude Fable 5.2 came out this week at $8 / $40 per million tokens, 20% below Fable 5.1 and twice the price of Opus 5.5. I wanted two answers: is it cheaper per fixed bug too, and does it fix more bugs than Opus?
I gave all three models the same 25 real bugs in Claude Code and scored every run with the maintainers' own tests.
- Fable 5.2 vs 5.1: 40% cheaper per run ($0.73 vs $1.22), twice the list-price cut. It wrote 27% fewer output tokens, read 67% less cache and finished in half the time (median 65 s vs 136 s).
- Fable 5.2 vs Opus 5.5: the same bugs. Both fixed 37 of 50 runs cleanly. The difference is 0 points (95% interval -8 to +8).
- Opus 5.5 was the cheapest fix: $0.54 per clean fix, against $0.99 for Fable 5.2 and $1.60 for Fable 5.1.
The setup
The 25 bugs are the ones from my earlier Haiku vs Sonnet vs Opus test: real issues fixed upstream between August and October 2026 in 13 open-source projects (js-yaml, jsdiff, qs, PapaParse, markdown-it, humanize, tomlkit, more-itertools, zod, cachetools, Python-Markdown, pyparsing, marked). 15 are "easy" (the cause sits near the symptom) and 10 are "hard" (the report only shows the symptom).
Each run starts at the commit before the fix, with the bug report cut down to symptom and expected result. Claude Code 2.1.322, headless, effort medium, tests allowed, no web, no installs. A run counts as a clean fix when every maintainer test for the fix passes and no other test breaks.
Pre-registered on GitHub before the first run. One dated amendment: the first 21 runs used my whole 5-hour subscription window in 34 minutes, so I cut the plan from 175 to 125 runs (Fable 5.2 × 2, Fable 5.1 × 1, Opus 5.5 × 2 per bug) before scoring anything.
Results
| Fable 5.2 | Fable 5.1 | Opus 5.5 | |
|---|---|---|---|
| Clean fixes | 37/50 (74%) | 19/25 (76%) | 37/50 (74%) |
| Cost per run | $0.73 | $1.22 | $0.40 |
| Cost per clean fix | $0.99 | $1.60 | $0.54 |
| Output tokens per run | 5,893 | 8,124 | 6,125 |
| Cache reads per run | 298K | 889K | 440K |
| Turns per run | 14.6 | 21.0 | 17.4 |
| Median time | 65 s | 136 s | 67 s |
Costs are API list prices from Claude Code's own log. On the easy bugs Fable 5.2 cost $0.48 a run against $0.26 for Opus; on the hard ones, $1.11 against $0.61.
All three models missed the same four bugs: a cachetools class-access case, a js-yaml BOM case, a tomlkit underscore case and a zod numeric-key case. None of them found what the report left out.
Limits
25 bugs, 2 runs per bug for Fable 5.2 and Opus, 1 for Fable 5.1, so Fable 5.1's fix rate is the least certain number here. One machine, one prompt style, effort medium only. These are bounded, single-issue bugs; harder multi-file work may separate the models where this didn't.
Data
All 135 runs (the 125 in the plan plus 10 finished before the amendment) with the full logs, diffs and scores, CC BY 4.0:
- Data page: https://tokensave.app/data/claude-fable-5-2-real-bugs
- GitHub (pre-registration, amendment, scripts): https://github.com/johnisking/token-cost-by-language/tree/main/claude-fable-5-2-real-bugs
