Benchmark Almanac

GDP.xlsx

Gemini 4 Argon 38.3% leads since September 2026
Not in model cards

Agents open a professional Excel workbook in a sandbox and answer a short question about it, as a domain professional would. A Gemini autograder checks the final answer against criteria.

Follows the format of GDP.pdf with spreadsheets. The full set has 70 Harbor tasks in 13 domains, and the HF dataset holds 3 samples.

Made by
Surge AI
Released
September 2026
Kind
Documents and spreadsheets
Level
Expert
Scored
Graded by a model
Answers
Agent task
Tasks
Kept private
Strongest model tested
GPT-6 Astra #3 on the capability index
Last model added
September 2026
Scores refreshed
November 12th, 2026
20
Models scored on its own leaderboard

Gemini 4 Argon leads with 38.3%

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best.

20 models. Scroll sideways for the rest.

  1. 38.3 Gemini 4 Argon
  2. 30.3 Claude Opus 5.5
  3. 29.1 Claude Sonnet 5.5
  4. 23.1 Claude Fable 5.1
  5. 22.9 GPT-6 Astra
  6. 22.6 Muse Spark 1.3
  7. 22.3 Grok 4.7
  8. 22.0 GPT-6.1 Sol
  9. 21.4 GLM-5.3
  10. 20.3 Qwen 3.8 Max
  11. 18.3 Tencent Hy4 preview
  12. 18.3 Gemini 3.8 Flash
  13. 17.1 GPT-6 Sol
  14. 16.9 Claude Sonnet 5
  15. 14.9 GPT-6 Luna
  16. 13.4 Kimi K3
  17. 10.3 Gemini 3.1 Pro
  18. 6.90 DeepSeek-V4-Pro
  19. 4.90 Nemotron 3 Ultra
  20. 0.0 Mistral Large 3
  1. 1 Gemini 4 Argon high 38.3%
  2. 2 Claude Opus 5.5 adaptive/max 30.3%
  3. 3 Claude Sonnet 5.5 adaptive/max 29.1%
  4. 4 Claude Fable 5.1 adaptive/max 23.1%
  5. 5 GPT-6 Astra max 22.9%
  6. 6 Muse Spark 1.3 max 22.6%
  7. 7 Grok 4.7 xhigh 22.3%
  8. 8 GPT-6.1 Sol max 22%
  9. 9 GLM-5.3 max 21.4%
  10. 10 Qwen 3.8 Max xhigh 20.3%
  11. 11 Tencent Hy4 preview high 18.3%
  12. 12 Gemini 3.8 Flash high 18.3%
  13. 13 GPT-6 Sol max 17.1%
  14. 14 Claude Sonnet 5 adaptive/max 16.9%
  15. 15 GPT-6 Luna max 14.9%
  16. 16 Kimi K3 max 13.4%
  17. 17 Gemini 3.1 Pro high 10.3%
  18. 18 DeepSeek-V4-Pro max 6.90%
  19. 19 Nemotron 3 Ultra 4.90%
  20. 20 Mistral Large 3 0%

Source: Official GDP.xlsx leaderboard

#ModelScoreSettingDateSource
1 Gemini 4 Argon 38.3% high Sep 2026 Official leaderboard
2 Claude Opus 5.5 30.3% adaptive/max Sep 2026 Official leaderboard
3 Claude Sonnet 5.5 29.1% adaptive/max Sep 2026 Official leaderboard
4 Claude Fable 5.1 23.1% adaptive/max Sep 2026 Official leaderboard
5 GPT-6 Astra 22.9% max Sep 2026 Official leaderboard
6 Muse Spark 1.3 22.6% max Sep 2026 Official leaderboard
7 Grok 4.7 22.3% xhigh Sep 2026 Official leaderboard
8 GPT-6.1 Sol 22% max Sep 2026 Official leaderboard
9 GLM-5.3 21.4% max Aug 2026 Official leaderboard
10 Qwen 3.8 Max 20.3% xhigh Aug 2026 Official leaderboard
11 Tencent Hy4 preview 18.3% high Aug 2026 Official leaderboard
12 Gemini 3.8 Flash 18.3% high Sep 2026 Official leaderboard
13 GPT-6 Sol 17.1% max Sep 2026 Official leaderboard
14 Claude Sonnet 5 16.9% adaptive/max Jun 2026 Official leaderboard
15 GPT-6 Luna 14.9% max Sep 2026 Official leaderboard
16 Kimi K3 13.4% max Jul 2026 Official leaderboard
17 Gemini 3.1 Pro 10.3% high Feb 2026 Official leaderboard
18 DeepSeek-V4-Pro 6.90% max Apr 2026 Official leaderboard
19 Nemotron 3 Ultra 4.90% Jun 2026 Official leaderboard
20 Mistral Large 3 0% Dec 2025 Official leaderboard

3 models have held number one on the official leaderboard since its release in September 2026

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic Google DeepMind
Own leaderboard 3 leaders
Gemini 4 Argon 38.3%

Source: Official GDP.xlsx leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Gemini 4 Argon Claude Opus 5.5 Claude Fable 5.1 Claude Sonnet 5 Gemini 3.1 Pro Mistral Large 3 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Gemini 4 Argon Claude Opus 5.5

Source: Official GDP.xlsx leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Economic work

All 97 in Economic work
  • JobBench In use 130 tasks from 35 occupations workers most want delegated. Agent works through messy reference files, graded by fact-anchored binary rubrics. GLM-5.3 61.4%
  • GDPval In use Real work tasks from 44 occupations with reference files. Experts blindly compare model deliverables with expert work. Metric is win-or-tie rate. Seed2.1 Pro 87.9%
  • GDPval-AA v2.1 In use Agents with shell and web access complete 220 GDPval gold tasks from 44 occupations. Blind pairwise comparisons give Elo ratings. Claude Opus 5.5 1,846
  • AA-Briefcase v1.1 In use Version 1.1 of Artificial Analysis's multi-hour office work evaluation, with Elo ratings fitted by Crowd-BT. Claude Opus 5.5 1,822
  • CoWorkBench In use Agent completes long-horizon productivity tasks across computer science, finance, law and medicine. Qwen in-house scoring (method unpublished). Claude Fable 5 75.9%
  • OfficeQA In use Grounded reasoning over a large corpus of U.S. Treasury Bulletin documents: locate tables and compute precise numeric answers. Claude Opus 4.7 86.3%