Mercor APEX-Accounting - Pass@1: leaderboard

APEX-Accounting: 100 professional month-end-close tasks across ten held-out company worlds, 1, 475 rubric criteria; a separate eleventh world is public. Native Loop-agent results. Pass@1 and mean criterion/rubric score remain separate metrics; no strongest-harness selection. Criterion-level Met/NotMet LM grading includes a scattergun penalty. The source states five-million-token/500-step limits, step retries and at most three failed-trajectory reruns. Its k=4 criterion means and separately described eight-run pass curves must not be treated as one identical repeated-trial population; native configuration remains retained. Exact native effort, API/EAP identity and score configuration remain per observation. Missing/failure and repeated-trial policies are not inferred from display sample counts. Metric: Pass@1 (%). Source: https://www.mercor.com/apex/apex-accounting-leaderboard/

Metric: Pass@1 (%). Source: www.mercor.com. Saturation forecast: Around 2031. 50 models tracked. Updated 17 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 4 ArgonHigh23.8#14
2Gemini 4 CarbonHigh22.5#3
3Claude Fable 5.2Max22.5#1
4Muse WonderMax21.3#31
5Gemini 4 HydrogenMercor Accounting run: source-id=gemini/gemini-4-hydrogen-eap, access=EAP20#13
6Claude Fable 5.1Max19.3#17
7Claude Sonnet 5.5Max19#21
8Claude Opus 5.5Max18.3#7
9Claude Opus 5Max17.8#23
10Kimi K3.1 ProMax17.2#32
11GPT-6.1 TerraMax16.8#22
12Muse Spark 1.3Max16.5#45
13Gemini 3.8 FlashHigh15.5#39
14Claude Haiku 5.5Max15.3#76
15Grok 4.5High14.5#91

No result here: #2 GPT-6.2 Astra, #6 GPT-6.2 Sol.

Interactive version: theaggregate.ai/benchmark?slug=mercor-apex-accounting-pass-1 · How It Works · Data refreshed daily, snapshot 2026-11-17.