Model results · Anthropic

Claude Fable 5.2

Every published VulcanBench measurement of Claude Fable 5.2, newest first: combined score, Code quality, runtime, tokens and list-price cost at every effort level, graded by deterministic hidden tests on real tasks. Measured through Claude Code on a Claude subscription; the refusal fallback never answered. Claude Fable 5.2 is a newer model than Fable 5.1, which has its own page.

Claude Fable 5.2 across every effort level

In Claude Code across five effort levels: combined score 91.92 at Low, 92.25 at Medium, 92.76 at High, 93.25 at Extra-high and 93.85 at Max, first on the Frontier v4 board. All 115 runs solved, every hidden test passed at every level; Code quality 80.63 to 86.02, judged by Muse Spark 1.3 and GPT-6.1 Sol; 4.3 to 37.1 minutes per task.

November 13, 2026 · 23 tasks · 5 effort levels · 115 runs of Claude Fable 5.2 · VulcanBench Frontier v4 · Code quality protocol v3.30.2 and v3.30.3

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

Claude Fable 5.2 leads at every level under the published scores, a matched Muse and Sol panel and Muse alone, by 0.08 at Medium under the matched panel. Its code quality is about level with Fable 5.1’s; its lead comes from correctness, speed and cost.

November 13, 2026 · comparison card · VulcanBench Frontier v4

Cost and tokens across effort levels

$1.72 to $10.88 per task at Claude Code’s list price, with output rising from 16.2K to 141.5K tokens a task, most of it thinking at the higher levels; the whole sweep prices at $509.85 for 115 runs.

November 13, 2026 · list-price estimates, solver inference only · VulcanBench Frontier v4

How Claude Fable 5.2 thinks

A deep look at its traces: it reads the reference binary before its first edit in every run, thinks in one long burst at Low, and spends its extra effort at Max on proving its rewrite.

November 12, 2026 · trace analysis · VulcanBench Frontier v4

Redacted agent traces: Claude Fable 5.2 on Frontier v4

The trace repository for the sweep in the report above, published separately on GitHub. Structure only: every tool call and file change with its timing, and thinking as lengths only, without the task content.

GitHub · vulcanbench-fable52-traces

Next

How would Claude Fable 5.2 do on your stack?

Public results answer "which model is stronger on this suite." Whether Claude Fable 5.2 belongs in your routing table depends on your languages, your codebases, and your task mix, that gets measured, not guessed. Grading rules are in the methodology.

Build an eval suite from your repo All reports