Fable 5.2's benchmark data comes in three batches: Anthropic's official 12-item table and system-card appendix, Artificial Analysis' independent measurement, and third-party boards such as Vals. All three batches are credible, but they use different harnesses, and mixing them produces wrong conclusions — Terminal-Bench 4.0's three numbers are the textbook case. This article lists every checkable score and flags each harness trap.
Artificial Analysis Intelligence Index v5.0.1 (listed Nov 10, 2026; model page read Nov 15): Max 65, No. 1 of 74 models; High/Xhigh 62, Medium 56, Low 54. The gains come mainly from scientific reasoning, knowledge and FrontierCode 2.0. It wins 5 of 9 evaluations, including HLE-Diamond 80% and Terminal-Bench-Science 73%. More first places: AA-3D (brief-to-3D scenes, human blind rating) at Elo 1271, and VIALS-AA (lab-image reading) at 67.1% (all-three-correct basis, No. 1 of 21).
But the same page states the price of 65: Max costs $26.54 per index task — the most expensive on the board, about 2.6× Opus 5.5 and about 4.8× Carbon; about 45.3 minutes per task (Nov 15 reading, drifting with AA re-runs; snapshots range 44.1–46.8 minutes), time to first token around 268 seconds; output speed at a 5-day average of about 97–103 tok/s, drifting as well. On VIALS-AA, Fable averaged 37.4 tool calls per task — the most of any Anthropic model, 4.7× Astra's. Much of the cost is the call volume.
Anthropic's launch-page benchmark table (each model at its highest tier unless noted; Fable/Opus with safeguards on; all vendor-reported): FrontierCode 2.0 49.1% (Xhigh; Opus 45.5%, Astra 42.8%, Carbon 41.1% at Medium), FrontierSWE v2 69.9% (5.1 56.3%), Terminal-Bench-Science 0.1 74.6% (Mythos 75.3%), Remote Labor Index 41.3% (High, 99 of 240 projects completed; Opus 21.3%), HLE (with tools) 74.2%, OSWorld 2.1 89.3%, WANDR 82.1%, Chartography 73.4%, RoboDojo Sim 64.9% (internal run). The system card adds SWE-bench Pro 93.0% (Max, mean of 5 trials; not in the launch-page table), SWE-bench Multilingual 96.6%, Multimodal 74.1%, CursorBench 4.0 60.0% (Max) / 56.9% (High), and HLE without tools 73.2%.
The system card says Fable 5.2 wins 14 of 16 headline results; the two losses are GDPval-AA and AA-Briefcase Elo — and the latter deserves a close read. Fable's rubric score of 66.1% beats Opus's 61.2%, but its Elo of 1785 sits below Opus's 1800 and Sonnet 5.5's 1813. Quote both numbers: Elo alone makes Fable look like the loser, the rubric alone makes it look unbeaten. The other loss is AutomationBench: Carbon 53.9%, Fable 51.4%.
Vals AI (Nov 9, run at High): Vals Index v2.2 at 71.31% ±0.96, No. 1 of 36 models; Vibe Code Bench v1.1 at 93.77%, No. 1 of 127; Finance Agent v2.1 69.47% and EMB v1.1 77.38%, both No. 1; Terminal-Bench 4.0 65.15%, tied No. 1 with Opus 5.5; $17.94 per test, 0% refusal rate. But the lead is inside the margin of error (Sonnet 5.5 70.85%, Carbon 70.45%), and weak sub-scores stand out: Tax Agent Bench 27/84 (32.09%), Legal Research Bench 19/93 (46.15%).
ARC Prize / Epoch AI (Nov 9–10, via secondary reports; the ARC-AGI-3 figures corrected Nov 18 against the ARC Prize Nov 17 board): ARC-AGI-2 93.8% (xhigh, ARC Prize verified, the highest verified score; max 92.5%, high 90.4%); ARC-AGI-3 81.2% (xhigh, best tier; high 79.3%, max 65.7%), against GPT-6.2 Astra's 70.3% (max) — this site had previously used the previous-generation GPT-6 Astra's score as the GPT-6.2 Astra comparison; GPQA Diamond 93.9% (Epoch AI, max tier). Xhigh beating max on ARC-AGI-2 matches the FrontierCode inversion. SimpleQA Verified at 94% (Epoch AI figures, strongest rival about 76%) is cited via the AIPings compilation.
Three numbers from three runners, not a typo: AA's independent run 59.6% (Max/High; Xhigh 61.6%), Anthropic's system-card self-test 67.6%, and Vals' run 65.15% (tied No. 1 with Opus 5.5). On AA's board, Fable's 59.6% ranks fourth (Kimi K3.1 Pro 67.2%, Carbon 65.2%, Sonnet 5.5 63.6% ahead), tied with Opus 5.5. Early coverage rounded Sulat's figure to "60%," and MultiAI's table carried the 67.6% system-card figure — this site's copy separates the three by runner; cite them the same way.
Two more easy confusions: AA's HLE-Diamond 80% and the vendor HLE 74.2% (with tools) / 73.2% (without tools) are different tests — don't mix them; Carbon's FrontierCode 41.1% was run at the Medium tier, so the conditions aren't fully equal. AA's time-per-task and output-speed figures drift with re-runs; this page uses the Nov 15, 2026 readings.
The weak spots, collected: Terminal-Bench 4.0's AA run at 59.6% ranks only fourth; AA-Briefcase Elo 1785 trails Opus 1800 and Sonnet 1813; AutomationBench 51.4% trails Carbon 53.9%; Vals' tax and legal sub-scores rank low (27/84, 19/93); per-task cost and time are the highest on the board; on FrontierCode 2.0, Max (45.7%) scores below Xhigh (49.1%), with 22 timeouts still counted. These don't invalidate the 65 — but they set Fable 5.2's boundaries: the hardest long-horizon tasks are worth its price; everyday tasks may not be.
Reference: Artificial Analysis — Claude Fable 5.2 model page (Index 65) (rewritten and organized by ModelBrief; facts per official sources and Artificial Analysis)