Argo-Bench: leaderboard

Argo-Bench (TextQL, 2026) — 210 enterprise data-science tasks on a simulated year of a New York food-delivery company exported to an Oracle E-Business Suite warehouse (235 tables, 7.5 billion rows). Agents query the warehouse, analyze in Python and file bans, forecasts, budgets and reported figures across trust and safety, marketplace, FP&A, accounting and growth; each filing is graded against the simulator's hidden state. Mean task score out of 100, one row per reasoning effort.

Metric: Mean task score (0-100) over 210 enterprise data-science tasks on a simulated Oracle EBS warehouse (v1.1, graded against the simulator's hidden state on a private-seed world); one row per reasoning effort. Source: argo-bench.com. Saturation forecast: Around June 2027. 21 models tracked. Updated 16 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Fable 5.2xHigh70.22#1
2Gemini 4 CarbonHigh62.35#2
3Claude Opus 5.5xHigh58.38#5
4Gemini 4 ArgonHigh53.86#12
5GPT-6 AstraxHigh50.69#6
6Claude Sonnet 5.5xHigh49.93#19
7GPT-6.1 SolxHigh47.81#10
8GPT-6.1 TerraxHigh44.21#21
9Muse WonderHigh41.42#28
10Grok 4.7xHigh38.84#47
11GPT-6 SolxHigh34.94#26
12DeepSeek V4.1 Flash (1017)xHigh31.79#41
13Kimi K3xHigh27.22#56
14Gemini 3.8 FlashMedium25.42#36
15DeepSeek V4.1 FlashMax23.21#62

Example tasks

We're planning merchant coverage for December. How many holiday closures should we expect?
Count closure events by their start date, including separate closures at the same merchant.
Forecast `december_holiday_closures`, with an 80% central prediction interval.

Source

Compliance thinks some of our storefronts are trading on a permit that belongs to a different business. Work out which ones and take them off the platform. Ban merchants with a call carrying the VENDOR_ID of every storefront you are deactivating.

Source

Interactive version: theaggregate.ai/benchmark?slug=argo-bench · How It Works · Data refreshed daily, snapshot 2026-11-16.