Fable 5.2 is 20% cheaper per token. On our bug fix it averaged $2.64 a run, under Fable 5.1's $3.28
Launch-day test: Claude Fable 5.2, Fable 5.1 and Opus 5.5 on the same planted bug in undici (902 files), same prompt and tools. Fable 5.2 (rerun on Claude Code 2.1.321) fixed the bug in 3 of 3 and averaged $2.64 a run and 2.37M tokens. Fable 5.1 averaged $3.28 and 2.04M; Opus $0.77 and 1.39M. Small samples: 3, 2 and 3 runs.
Other checks: Fable 5 2 run 3 missed “ran the tests after the last code edit (TRACE.md edits don't count)”; Opus 5 5 runs 2 and 3 missed “ran the tests after the last code edit (TRACE.md edits don't count)”; Opus 5 5 run 3 missed “existing test/max-response-size.js passes”. Fable 5.2 (rerun) fixed the bug in 3 of 3 and passed all six checks in 2 of 3. Fable 5.1 passed all three listed checks in 2 of 2. Opus passed them in 2 of 3; its third run hit the 15-minute limit with its fix temporarily reverted.
Anthropic released Claude Fable 5.2 today, at $8 per million input tokens and $40 per million output, down from $10 and $50 for Fable 5.1 (pricing). A lower price per token doesn’t tell you what a task costs, so we ran it on a real repo the same afternoon, next to Fable 5.1 and Opus 5.5.
Update Nov 9
The Fable 5.2 numbers in the table below are from an evening rerun on Claude Code 2.1.321, which recognizes claude-fable-5-2 (list pricing, 1M context, 128K max output). Our first Fable 5.2 pass that afternoon used Claude Code 2.1.312, which did not recognize the model: it treated it as a 200K context window and 32K max output, reported cost basis “unknown,” and one of those two runs hit our 15-minute limit. Those first-pass figures ($4.40 avg, 4.37M tokens, fixed 2/2, all six checks 1/2) are kept below for reference only. The rerun was graded on Node 24.11, the same Node as the original test. Fable 5.1 and Opus 5.5 numbers are unchanged from the afternoon runs on 2.1.312.
The setup
Same task as our Nov 4 test: undici, Node’s HTTP client, at commit 6228fc5 (902 files), with one planted bug: maxResponseSize is checked per chunk, not on the running total, so a big body in small chunks gets through. The prompt asks Claude to trace the option through the code into TRACE.md, check five other ways people set it, fix the bug, add a regression test and run the tests.
Three models, nothing else changed: claude-fable-5-2, claude-fable-5-1 and claude-opus-5-5, each at its default effort, in headless Claude Code, with the same prompt, tools, 60-turn cap and 15-minute limit per run. Fable 5.1 and Opus ran on Claude Code 2.1.312. The Fable 5.2 column is the 2.1.321 rerun.
Six checks graded each run afterwards, including a hidden 7-case test Claude never saw.
Runs: 3 for Fable 5.2 (rerun), 2 for Fable 5.1, 3 for Opus 5.5. The afternoon Fable passes were cut to 2 each to stay under a $24 slot; the evening Fable 5.2 rerun had its own $13 cap and finished three runs for $8.00 including the recognition smoke.
What happened
| Fable 5.2 | Fable 5.1 | Opus 5.5 | |
|---|---|---|---|
| Runs | 3 (rerun, CC 2.1.321) | 2 | 3 |
| Fixed the bug (hidden test, 7 cases) | 3 of 3 | 2 of 2 | 2 of 3 |
| Existing max-response-size test passes | 3 of 3 | 2 of 2 | 2 of 3 |
| TRACE.md named all 10 files | 3 of 3 | 2 of 2 | 2 of 3 |
| Ran the tests after the last code edit | 2 of 3 | 2 of 2 | 1 of 3 |
| Passed all 6 checks | 2 of 3 | 2 of 2 | 1 of 3 |
| Hit the 15-minute limit | 0 of 3 | 0 of 2 | 1 of 3 |
| Average cost per run | $2.64 | $3.28 | $0.77 |
| Cost range | $2.49 to $2.93 | $2.53 to $4.02 | $0.64 to $0.88 |
| Average tokens per run | 2,374,934 | 2,040,607 | 1,390,067 |
| Average output tokens | 27,610 | 33,223 | 13,892 |
| Largest main-session context, average | 97,777 | 103,315 | 58,507 |
| Wall time per run | 12:29, 8:11, 7:17 | 11:10, 7:22 | 2:25, 2:41, 15:00 (limit) |
Claude Code claude -p (2.1.321 for Fable 5.2; 2.1.312 for Fable 5.1 and Opus). Token counts include cached input. No run in any setup started a subagent. Every run, check and diff for the table above (Fable 5.2 rerun + Fable 5.1 + Opus): run data. The afternoon-only Fable 5.2 pass on Claude Code 2.1.312 remains in that folder’s earlier merge for reference, and the standalone rerun folder is runner/results/fable-5-2-rerun-cc2.1.321/.
First Fable 5.2 pass (CC 2.1.312, for reference): 2 runs, fixed the bug 2 of 2, passed all 6 checks 1 of 2, hit the 15-minute limit 1 of 2, averaged $4.40 a run ($3.98 to $4.82) and 4,372,293 tokens. Claude Code did not recognize the model yet, so we priced those two runs from the published list rates.
What we learned
Once Claude Code recognized Fable 5.2, the lower per-token price did make this task cheaper than Fable 5.1. The rerun averaged $2.64 a run against Fable 5.1’s $3.28 (about 19% less), with 2.37M tokens vs 2.04M and a smaller peak main-session context (97,777 vs 103,315). That is a different story from the first pass under 2.1.312, which averaged $4.40 and 4.37M tokens with a 157,682 peak context.
Fable 5.2 fixed the bug in every scored rerun. All three runs applied the one-line running-total fix. Two of three passed all six checks; the third missed re-running tests after the last code edit. None hit the 15-minute limit.
Opus 5.5 was still the bargain, with one bad run. Two Opus runs fixed the bug in under 3 minutes at $0.79 and $0.88. The third ran into the 15-minute limit while it had temporarily put the bug back to check its new test, so it failed 5 of 6 checks. On Nov 4, Opus passed this task in 10 of 10 runs.
What failed, and the limits of this test
- Tiny samples. 3, 2 and 3 runs. One more bad run in any column would change the story, so read this as a first look, not a ranking.
- The first Fable 5.2 client was wrong. Claude Code 2.1.312 reported cost basis “unknown” and assumed 200K context / 32K max output. We kept those two runs for reference and replaced the Fable 5.2 column with the 2.1.321 rerun. A client that doesn’t know the model can inflate tokens and wall time.
- One Opus run was cut off at 15 minutes, so Claude Code never reported its cost. We added up the token usage from Claude Code’s own session logs instead. That method matched the reported cost of finished runs to the sixth decimal, but a request still in flight when the run stopped isn’t counted, so that $0.64 figure is a lower bound.
- One task, one repo, and the bug is planted. Harder, open-ended work is where Anthropic pitches Fable; a one-line bug fix with a trace may not show what it’s for.
- API cost only. Afternoon slot $17.84 (7 runs + model checks) plus evening Fable 5.2 rerun $8.00 (3 runs + recognition smoke) is about $25.84. We didn’t measure Pro or Max plan usage limits.
The recipe
- Don’t pick a model from the price sheet alone. On this task Fable 5.2’s lower per-token price only showed up as a lower per-run cost after Claude Code recognized the model.
- For a trace-and-fix like this, start with Opus 5.5. Two of three runs fixed the bug at under $0.90 (one of those missed re-running tests after a TRACE edit; only 1 of 3 passed all six checks); budget for a retry.
- If you try Fable 5.2, use a Claude Code that lists it and check the cost basis. 2.1.321 reports list pricing with a 1M context window and 128K max output. 2.1.312 did not.
- Measure on your own repo. Run the same prompt two or three times per model and compare cost per passing run, not cost per token.