Claude Fable 5.2 across every effort level

Across every effort level

Claude Fable 5.2 ran every task once at each of the five effort levels Claude Code offers for it (Extra-high is Claude Code’s xhigh), 23 tasks per cell, in Claude Code 2.1.321 on the owner’s Mac, one task at a time. Low to Extra-high ran on November 9 and 10; Max waited about a day for the Fable usage limit to reset and ran on November 10 and 11. No run reached the flat 3-hour task bound; the longest, Max reflowcore, took 52.1 minutes. All 115 runs passed every hidden test and carry a published Code quality score.

Under the prior 50/15/15/20 profile applied to the same Code quality scores the ladder has the same shape, 93.14 to 94.48. The card and the full per-effort table follow.

Beside Claude Fable 5.1, Claude Opus 5.5 and GPT-6 Astra

On the published scores, the ones the board uses, Claude Fable 5.2 is ahead of all three at every level, and its Low level (91.92) is above every Fable 5.1, Opus 5.5 and Astra cell (the highest, Fable 5.1 at Max, is 91.84). The panels differ, though: Fable 5.2 is judged by Muse Spark 1.3 and GPT-6.1 Sol, the other three by Muse Spark 1.3 and Grok 4.6, and GPT-6.1 Sol rates Fable 5.2’s code about 7 points higher than Muse does. The next section removes that difference two ways.

Speed and cost are where the gap is largest. At Low Fable 5.2 costs $1.72 a task, about the same as GPT-6 Astra ($1.72) and Opus 5.5 ($1.70), and takes 4.6 minutes against 4.2 and 13.7, while solving 23 tasks to their 22 and 17. Fable 5.1 took 26.9 to 38.8 minutes a task at every level and cost $7.96 to $14.49. At Max Fable 5.2 is the most expensive of the four, $10.88 against $9.06, $8.75 and $2.57.

Matched panel and Muse alone

Matched panel. For the card, GPT-6.1 Sol also reviewed the frozen Claude Fable 5.1 and Claude Opus 5.5 patches beside Muse’s existing reviews, in comparison-only rounds v3.30.4 and v3.30.5. These figures are for this comparison only; the board keeps Fable 5.1’s and Opus 5.5’s published Muse and Grok scores. Under the matched panel Fable 5.2 still leads at every level, but only narrowly at Medium, where it is 0.08 ahead of Opus 5.5, well inside one standard error of the difference (0.72). Fable 5.1 at Max (92.69) is ahead of Fable 5.2 at Low (91.92), and Opus 5.5’s best level (Medium, 92.16) is 0.24 above it. GPT-6 Astra cannot join this panel, because GPT-6.1 Sol shares its lab.

Sol reviews every model higher than Muse does, and by different amounts: 6.7 to 8.1 points for Fable 5.2, 3.1 to 5.5 for Fable 5.1 and 5.8 to 10.1 for Opus 5.5. Under the matched panel Fable 5.2’s Code quality is within 1.5 points of Fable 5.1’s at every level, below it from Low to Extra-high and 1.0 above it at Max; every gap is inside the standard error of the difference (1.6 to 2.1).

Muse alone. Muse Spark 1.3 sat on every one of these rounds, Astra’s included, with the same pinned binary, rubric and weights, so its scores alone are the only like-for-like comparison that includes all four models.

Judged by Muse alone, Fable 5.2’s code is slightly below Fable 5.1’s at every level, by 0.7 to 2.0 points, each smaller than the standard error of the difference (1.8 to 2.6). It still leads on combined score at every level, because it solves every task. Under both panels its lead comes from correctness, speed and cost, not from code quality.

How it works, and how that differs from Fable 5.1

Fable 5.2’s runs recorded summaries of its thinking (a run condition earlier Claude columns did not have) and every tool call. They are withheld because they describe the reference programs, but their structure is public in the session traces. Fable 5.1’s raw runs did not survive a host reset in September, so the comparison with it uses what both left in the same form: usage receipts, durations, fallback records and patches.

• Effort controls how much it thinks. From Low to Max its output per task grows 8.7 times (16 thousand to 142 thousand tokens) and thinking rises from 48% to 67% of it. Fable 5.1’s output grew only 1.2 times, and its time stayed between 27 and 39 minutes at every level.
• It is lean at low effort. At Low it used 19% of Fable 5.1’s output tokens, 17% of its time and 22% of its cost, and solved 23 tasks to Fable 5.1’s 19.
• The patch does not grow. In every run it changed exactly one file, the module under repair, with a median of 97 to 113 added lines at every level, and kept its test harnesses in scratch space. Fable 5.1’s median patch grew from 112 to 190 lines with effort and added test files in 5 to 11 runs per level.
• It reads the binary first. It inspected the reference binary in all 115 runs, at a median of its second tool call at Low and its seventh at Max, and the extra effort at higher levels goes to checking its rewrite against the binary rather than to a different answer.

Report 01, How Claude Fable 5.2 thinks, follows these patterns through its traces in depth, with comparisons to GPT-6 Astra and Claude Opus 5.5.

Memorization check

Every Frontier v4 reference fix and test has been public on GitHub since late August 2026, about ten weeks before Claude Fable 5.2, and it solved every run, so the harness checked whether it was reciting answers.

• Copying. Its solved patches reproduce 5.6% of the reference fixes’ distinctive lines verbatim, in the same range as six other models (4.4% to 5.7%), including Claude Opus 5, which was benchmarked before the fixes were public. A recited fix would score near 100%.
• Working it out. It inspected the legacy binary before its first edit in 115 of 115 runs, against 43% to 67% for Claude Opus 5, Sonnet 5.5 and Haiku 5.5.
• Recall. With no tools, asked to reproduce five reference fixes from memory, it answered each time that it did not remember the fix and wrote none; asked to fix one task blind, it declined to write a diff.

Cost and tokens across effort levels

All 115 runs are priced at Claude Code’s own list-price total for each session, the cost basis of every Claude column, with judging excluded. These are list-price estimates, not subscription charges: Claude Fable 5.2 ran on a Claude subscription. Cost per task is $1.72 at Low and Medium, $2.86 at High, $4.99 at Extra-high and $10.88 at Max; the whole sweep comes to $509.85. The extra cost is extra thinking: mean output rises from 16.2K tokens a task at Low to 141.5K at Max, of which thinking is 7.7K and 94.7K, and tool calls from about 11 to 59 a task.

On the Frontier v4 board

Claude Fable 5.2 takes first place on the board’s 86 model and effort columns: Max ranks 1st (93.85 ± 0.42, n=23), Extra-high 2nd, High 4th, Medium 7th and Low 10th; Grok 4.7 at Extra-high (93.15) is third. Max carries the best tag. Its spread from Low to Max is under 2 points, so the board’s effort suggestions pick Extra-high for critical work and Medium for routine and rough work. The leaderboard carries every column; Claude Fable 5.2’s cells carry the ‖‖ footnote with the panel disclosure on this page.

Read these placements with the panel in mind. Claude Fable 5.2 is judged by Muse Spark 1.3 and GPT-6.1 Sol, as are Grok 4.7, DeepSeek V4.1 Flash and Reflection Beam; most other columns, Claude Fable 5.1 and Claude Opus 5.5 included, by Muse Spark 1.3 and Grok 4.6. Scores from different panels are not calibrated against each other, so gaps of a point or two between columns judged by different panels are not a reliable order. By Muse alone, Fable 5.2’s combined score is 91.09 to 92.78.

What Code quality measures

Every judge receives the same rubric, frozen by hash before any review. It names the reader it scores for: an engineer who has never seen the code, reads it top to bottom without running it, and must make a correct change in one sitting. It tells the judge that its own ease at parsing dense code is not evidence of readability. Six dimensions are scored 0 to 4 in half steps against written anchors: naming, presentation and intent form the human readability sub-score; structure, changeability and verifiability form the maintainability sub-score. Every score must cite an exact excerpt from the code and a concrete consequence for that reader. The host computes the sub-scores; the judge does no arithmetic.

The 33 points have three designed layers. The reviewed panel carries 24 points. An intent-recovery probe carries 9: the judge, given only the specification and the code, lists where the code departs from the specification, and a separate call matches that list against a frozen answer key, over the departures the submission passed tests for. The third layer, measured maintenance, is designed for 12 points but is not yet built, so the pre-registered 24 plus 9 split is in force.

Code quality rises at every level, by 5.40 points from Low to Max. Human readability is the weaker sub-score at every level. GPT-6.1 Sol is the higher rater at every level, by 6.7 to 8.1 points on the reviewed score. Per-run sub-scores from each judge are in runs.csv; the exact rubric, system text, probe and match instructions are in judge-protocols.json.

Judges and calibration

Round v3.30 seated Muse Spark 1.3 (Meta) and Grok 4.6 (xAI), the panel of the Claude Opus 5.5 column, and both retook the calibration exam. Grok 4.6 failed one gate, g14_pair_2_1 (a pair of controls it scored 100 one way round and 50 the other), which the one-gate allowance does not cover, so it judged nothing. Muse Spark 1.3 passed every gate and judged Low to Extra-high under v3.30 and Max under v3.30.1. On the owner’s choice, GPT-6.1 Sol (OpenAI) through the Codex CLI at medium effort took Grok’s seat, passed every gate under v3.30.2, and judged Low to Extra-high under v3.30.2 and Max under v3.30.3, on byte-copied freezes of the same populations. Neither judge’s lab is Anthropic. The rubric, prompts, schemas, seed, repeats, gate allowance, weights and both judges’ pins equal the published v3.27 record, the Muse and Sol round that judged DeepSeek V4.1 Flash.

Before scoring a single submission, each judge reviews ten held-out programs that implement the same ledger specification, each five times in a seeded order, and twenty gates fixed in advance check that the judge sees the construct; a judge may miss at most one gate by at most half a point.

Every gate value and control mean is in calibration.json.

One judge alone. Scored by one judge alone, the combined score would be (Muse or Sol) 91.09 or 92.74 at Low; 91.16 or 93.34 at Medium; 91.53 or 93.99 at High; 92.13 or 94.37 at Extra-high; 92.78 or 94.91 at Max. Either way the ladder rises at every level.

Counted calls and recoveries. GPT-6.1 Sol made 80 calibration calls and 92 primary reviews, 4 repeats, 8 pairwise checks, 92 intent probes and 92 answer-key matches under v3.30.2, and 23 primary reviews, 1 repeat, 23 probes and 23 matches under v3.30.3. Muse Spark 1.3 made the same counts under v3.30 and v3.30.1. Twelve of Sol’s answer-key matches (8 and 4) numbered the departures from one instead of zero and were published through the protocol’s frozen rule for that slip, recover_one_based_indexes, which leaves every score unchanged because intent scoring reads match statuses only. One Muse intent probe under v3.30 quoted an excerpt with its line breaks changed; it was re-wrapped to the source’s line breaks, after which it had to be verbatim in the code and pass the frozen validator. No other call was edited, there is no operator marker, and neither judge produced a reviewer fallback. In UTC, Muse judged on November 10 (Low to Extra-high) and November 12 (Max), and Sol on November 12. Every operator record is in judge-protocols.json.

How it was run

• The model. Claude Fable 5.2 as claude-fable-5-2 in Claude Code 2.1.321, pinned by hash with auto-updates off: the first release that names the model. Claude Code reports the serving model on every reply, and every reply in every run is claude-fable-5-2.
• Harness settings. Permission mode auto, safe mode, no web access, a Claude subscription, refusal fallback on as for every Claude column (never used). Every integrity audit is clean: no web access, no benchmark data or answer-key paths. Every file written outside the workspace was scratch space under the temporary directory.
• Thinking summaries recorded. Runs used --thinking-display summarized, so the traces keep summaries of the model’s thinking, a run condition earlier Claude columns did not have. The summaries are withheld from the public traces, which keep thinking as lengths only, because they describe the reference programs.
• Five levels. Claude Code takes low, medium, high, xhigh and max for Claude Fable 5.2; the board’s Extra-high is xhigh.
• Max waited for the usage limit. The Fable weekly usage limit was used up by the end of Extra-high. The first Max attempts were refused by the limit, did no work, hold no summary and are kept apart, not part of the population. Max ran from 23:15 PST on November 10, when the limit reset, to 13:28 PST on November 11.
• From a tagged run worktree. Every run ran from bench/2026-11-09-fable52-frontier-v4, so every run summary carries a clean source block and records the frozen suite lock’s task hash.
• Two receipts in two runs. Claude Code 2.1.321 can emit a second result receipt when a background task resumes a session after the model stops. It happened in two runs; their totals sum both receipts, as described under cost.
• Where and how. The owner’s Mac, one task at a time, on November 9 to 11, 2026. Runs were not repeated, and effort labels are Claude Code’s own.

Evidence and reproduction

Public files contain all 115 run measurements, both judges’ six-dimension sub-scores and intent-recovery scores per run, the combined score from each judge alone, hidden behaviours fixed, integrity-audit verdicts, each run’s task hash, raw and thinking tokens, tool-call counts and list-price cost per run, five aggregates, every calibration gate value (Grok’s failed exam included), the exact judge protocol text, every operator recovery, the card’s per-level table, the Muse-alone table and source hashes. Each run’s Code quality is (0.24 × reviewed score + 0.09 × intent recovery) / 0.33, where both terms are the equal mean of the two judges. The combined score is 100 × (0.50F + 0.085Q + 0.085S + 0.33C / 100). The reproduction guide gives the public arithmetic checks and links the protocol document, the operations log, the decision record, the Fable 5.2 results and the code in the harness repository.

No human rated anything; the scores are model judgment for a human reader, not human validation. Hash checks bind the export to frozen files; they do not prove that judges were unbiased or that no training overlap exists.

Withheld: raw prompts and responses, thinking summaries, submitted patches and reconstructed sources, the quirk answer keys (they describe hidden-test behaviour), reviewer session identifiers and usage receipts, solver event streams (a structure-only redacted trace of every run, with every tool call and its timing and thinking as lengths only, is public in vulcanbench-fable52-traces), and host binary paths. Summed solver time is 26.07 hours; judging is excluded.