Benchmark Almanac

APEX-Agents

Gemini 4 Argon 82.4% leads since Sep 2026 In use

480 long-horizon tasks in banking, consulting and law worlds with docs, sheets, email and chat. Rubric-graded, Pass@1.

Repo is Mercor's Archipelago harness that runs the dataset. Dataset license forbids training on it.

In the authors’ words

We introduce the AI Productivity Index for Agents (APEX-Agents), a benchmark for assessing whether AI agents can execute long-horizon, cross-application tasks created by investment banking analysts, management consultants, and corporate lawyers. APEX-Agents requires agents to navigate realistic work environments with files and tools. We test eight agents for the leaderboard using Pass@1. Gemini 3 Flash (Thinking=High) achieves the highest score of 24.0%, followed by GPT-5.2 (Thinking=High), Claude Opus 4.5 (Thinking=High), and Gemini 3 Pro (Thinking=High). We open source the APEX-Agents benchmark (n=480) with all prompts, rubrics, gold outputs, files, and metadata. We also open source Archipelago, our infrastructure for agent execution and evaluation.

Abstract of APEX-Agents (arXiv:2601.14242, CC0)
Made by Mercor Released Jan 2026 Simulated companies All tasks public Last in a launch post Oct 2026 Last model added Oct 2026 Scores refreshed 31 Oct 2026
Level
Expert Specialists with years of training, often over hours of work
Scored
Graded by a model A language model scores each response, usually against a rubric.
Answers
Agent task Many steps in a terminal, browser, desktop, app, game or robot.
Mercor-Intelligence/archipelago 296 stars Paper Dataset Leaderboard mercor.com
56
Models scored in Epoch AI data
11
Launch posts list it, from 7 labs
5
System cards report it, from 5 labs

Gemini 4 Argon leads with 82.4%

Best score per model in Epoch AI’s data. Where a model ran with several settings or harnesses, we keep its best.

56 models. Scroll sideways for the rest.

  1. 82.4 Gemini 4 Argon
  2. 75.5 Claude Sonnet 5.5
  3. 73.5 Claude Opus 5.5
  4. 69.8 DeepSeek V4.1 Flash 1017
  5. 68.6 Claude Fable 5.1
  6. 68.3 GPT-6.1 Terra
  7. 67.8 Gemini 3.7 Flash
  8. 66.1 Muse Wonder
  9. 65.8 Claude Opus 5
  10. 65.3 Grok 4.6
  11. 64.7 GPT-6 Astra
  12. 64.3 Gemini 3.8 Flash
  13. 63.6 Claude Fable 5
  14. 63.3 Qwen 3.8 Max
  15. 60.0 GPT-6.1 Sol
  16. 58.6 Muse Spark 1.3
  17. 58.2 GPT-5.6 Terra
  18. 56.6 GLM-5.3
  19. 56.2 Grok 4.5
  20. 56.1 Claude Haiku 5.5
  21. 55.1 GPT-5.5
  22. 54.6 Grok 4.7
  23. 54.5 Claude Sonnet 5
  24. 54.4 MiMo-V2.6 Pro
  25. 54.3 GPT-6 Sol
  26. 52.8 GLM-5.3-Flash
  27. 52.4 GPT-5.4
  28. 51.6 Mistral Large 4
  29. 51.4 GPT-5.6 Sol
  30. 49.2 Claude Opus 4.7
  31. 48.9 Claude Opus 4.8
  32. 48.8 Beam 1.0
  33. 47.8 Kimi K3
  34. 47.5 Qwen 3.8 27B
  35. 47.3 DeepSeek V4 Pro 0813
  36. 46.9 Gemini 3.6 Flash
  37. 46.3 Claude Opus 4.6
  38. 45.4 Muse Spark 1.2
  39. 45.2 GLM-5.2
  40. 44.3 GPT-6 Luna
  41. 43.0 Claude Sonnet 4.6
  42. 43.0 GPT-5.6 Luna
  43. 40.9 GLM-5.1
  44. 39.5 DeepSeek V4.1 Flash
  45. 37.7 MiniMax-M3
  46. 37.6 Kimi K2.7 Code
  47. 35.1 Gemini 3.1 Pro
  48. 33.8 Inkling
  49. 33.4 Muse Spark Open
  50. 31.8 Muse Spark 1.1
  51. 29.3 Gemini 3.5 Flash-Lite
  52. 27.5 Gemini 3.5 Flash
  53. 24.9 Qwen3.5 397B-A17B
  54. 22.7 Nemotron 3 Ultra
  55. 21.3 DeepSeek-V3.2
  56. 4.40 gpt-oss-120b
  1. 1 Gemini 4 Argon 82.4%
  2. 2 Claude Sonnet 5.5 max 75.5%
  3. 3 Claude Opus 5.5 max 73.5%
  4. 4 DeepSeek V4.1 Flash 1017 max 69.8%
  5. 5 Claude Fable 5.1 68.6%
  6. 6 GPT-6.1 Terra max 68.3%
  7. 7 Gemini 3.7 Flash 67.8%
  8. 8 Muse Wonder max 66.1%
  9. 9 Claude Opus 5 max 65.8%
  10. 10 Grok 4.6 65.3%
  11. 11 GPT-6 Astra unknown thinking 64.7%
  12. 12 Gemini 3.8 Flash 64.3%
  13. 13 Claude Fable 5 63.6%
  14. 14 Qwen 3.8 Max 63.3%
  15. 15 GPT-6.1 Sol max 60%
  16. 16 Muse Spark 1.3 58.6%
  17. 17 GPT-5.6 Terra max 58.2%
  18. 18 GLM-5.3 unknown thinking 56.6%
  19. 19 Grok 4.5 unknown thinking 56.2%
  20. 20 Claude Haiku 5.5 max 56.1%
  21. 21 GPT-5.5 unknown thinking 55.1%
  22. 22 Grok 4.7 54.6%
  23. 23 Claude Sonnet 5 unknown thinking 54.5%
  24. 24 MiMo-V2.6 Pro 54.4%
  25. 25 GPT-6 Sol max 54.3%
  26. 26 GLM-5.3-Flash 52.8%
  27. 27 GPT-5.4 unknown thinking 52.4%
  28. 28 Mistral Large 4 high 51.6%
  29. 29 GPT-5.6 Sol pro, max 51.4%
  30. 30 Claude Opus 4.7 max 49.2%
  31. 31 Claude Opus 4.8 max 48.9%
  32. 32 Beam 1.0 max 48.8%
  33. 33 Kimi K3 47.8%
  34. 34 Qwen 3.8 27B 47.5%
  35. 35 DeepSeek V4 Pro 0813 47.3%
  36. 36 Gemini 3.6 Flash 46.9%
  37. 37 Claude Opus 4.6 max 46.3%
  38. 38 Muse Spark 1.2 45.4%
  39. 39 GLM-5.2 45.2%
  40. 40 GPT-6 Luna max 44.3%
  41. 41 Claude Sonnet 4.6 high 43%
  42. 42 GPT-5.6 Luna unknown thinking 43%
  43. 43 GLM-5.1 40.9%
  44. 44 DeepSeek V4.1 Flash 39.5%
  45. 45 MiniMax-M3 37.7%
  46. 46 Kimi K2.7 Code 37.6%
  47. 47 Gemini 3.1 Pro 35.1%
  48. 48 Inkling 33.8%
  49. 49 Muse Spark Open 33.4%
  50. 50 Muse Spark 1.1 31.8%
  51. 51 Gemini 3.5 Flash-Lite 29.3%
  52. 52 Gemini 3.5 Flash 27.5%
  53. 53 Qwen3.5 397B-A17B 24.9%
  54. 54 Nemotron 3 Ultra 22.7%
  55. 55 DeepSeek-V3.2 Thinking; Novita 21.3%
  56. 56 gpt-oss-120b 4.40%
#ModelScoreSettingDateSource
1 Gemini 4 Argon 82.4% Sep 2026 Epoch AI
2 Claude Sonnet 5.5 75.5% max Sep 2026 Epoch AI
3 Claude Opus 5.5 73.5% max Sep 2026 Epoch AI
4 DeepSeek V4.1 Flash 1017 69.8% max Oct 2026 Epoch AI
5 Claude Fable 5.1 68.6% Sep 2026 Epoch AI
6 GPT-6.1 Terra 68.3% max Oct 2026 Epoch AI
7 Gemini 3.7 Flash 67.8% Aug 2026 Epoch AI
8 Muse Wonder 66.1% max Oct 2026 Epoch AI
9 Claude Opus 5 65.8% max Jul 2026 Epoch AI
10 Grok 4.6 65.3% Aug 2026 Epoch AI
11 GPT-6 Astra 64.7% unknown thinking Sep 2026 Epoch AI
12 Gemini 3.8 Flash 64.3% Sep 2026 Epoch AI
13 Claude Fable 5 63.6% Jun 2026 Epoch AI
14 Qwen 3.8 Max 63.3% Aug 2026 Epoch AI
15 GPT-6.1 Sol 60% max Sep 2026 Epoch AI
16 Muse Spark 1.3 58.6% Sep 2026 Epoch AI
17 GPT-5.6 Terra 58.2% max Jul 2026 Epoch AI
18 GLM-5.3 56.6% unknown thinking Aug 2026 Epoch AI
19 Grok 4.5 56.2% unknown thinking Jul 2026 Epoch AI
20 Claude Haiku 5.5 56.1% max Oct 2026 Epoch AI
21 GPT-5.5 55.1% unknown thinking Apr 2026 Epoch AI
22 Grok 4.7 54.6% Sep 2026 Epoch AI
23 Claude Sonnet 5 54.5% unknown thinking Jun 2026 Epoch AI
24 MiMo-V2.6 Pro 54.4% Sep 2026 Epoch AI
25 GPT-6 Sol 54.3% max Sep 2026 Epoch AI
26 GLM-5.3-Flash 52.8% Aug 2026 Epoch AI
27 GPT-5.4 52.4% unknown thinking Mar 2026 Epoch AI
28 Mistral Large 4 51.6% high Oct 2026 Epoch AI
29 GPT-5.6 Sol 51.4% pro, max Jul 2026 Epoch AI
30 Claude Opus 4.7 49.2% max Apr 2026 Epoch AI
31 Claude Opus 4.8 48.9% max May 2026 Epoch AI
32 Beam 1.0 48.8% max Oct 2026 Epoch AI
33 Kimi K3 47.8% Jul 2026 Epoch AI
34 Qwen 3.8 27B 47.5% Aug 2026 Epoch AI
35 DeepSeek V4 Pro 0813 47.3% Aug 2026 Epoch AI
36 Gemini 3.6 Flash 46.9% Jul 2026 Epoch AI
37 Claude Opus 4.6 46.3% max Feb 2026 Epoch AI
38 Muse Spark 1.2 45.4% Aug 2026 Epoch AI
39 GLM-5.2 45.2% Jun 2026 Epoch AI
40 GPT-6 Luna 44.3% max Sep 2026 Epoch AI
41 Claude Sonnet 4.6 43% high Feb 2026 Epoch AI
42 GPT-5.6 Luna 43% unknown thinking Jul 2026 Epoch AI
43 GLM-5.1 40.9% Apr 2026 Epoch AI
44 DeepSeek V4.1 Flash 39.5% Sep 2026 Epoch AI
45 MiniMax-M3 37.7% Jun 2026 Epoch AI
46 Kimi K2.7 Code 37.6% Jun 2026 Epoch AI
47 Gemini 3.1 Pro 35.1% Feb 2026 Epoch AI
48 Inkling 33.8% Jul 2026 Epoch AI
49 Muse Spark Open 33.4% Oct 2026 Epoch AI
50 Muse Spark 1.1 31.8% Jul 2026 Epoch AI
51 Gemini 3.5 Flash-Lite 29.3% Jul 2026 Epoch AI
52 Gemini 3.5 Flash 27.5% May 2026 Epoch AI
53 Qwen3.5 397B-A17B 24.9% Feb 2026 Epoch AI
54 Nemotron 3 Ultra 22.7% Jun 2026 Epoch AI
55 DeepSeek-V3.2 21.3% Thinking; Novita Dec 2025 Epoch AI
56 gpt-oss-120b 4.40% Aug 2025 Epoch AI

Labs report scores for 12 models in their cards. We plot every source in the charts below.

11 models have held number one since its release in Jan 2026

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for Epoch AI’s runs and what labs claim in their cards.

DeepSeek Anthropic OpenAI Google DeepMind ByteDance Seed Moonshot AI xAI
Epoch AI runs 11 leaders
Gemini 4 Argon 82.4%
Lab claims 5 leaders
Kimi K3.1 Pro 67.2%

Scores over time

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

Run or collected by Epoch AI Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Gemini 4 Argon Claude Sonnet 5.5 Claude Fable 5.1 Claude Fable 5 GPT-5.5 Claude Opus 4.6 DeepSeek-V3.2 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Gemini 4 Argon Claude Sonnet 5.5

6 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

Google DeepMind
Moonshot AI
MiniMax
Tencent
Upstage
xAI
StepFun
DateLaunch postScoreAs the post writes it
24 Oct 2026 Kimi K3.1 Pro 67.2 APEX-Agents
20 Sep 2026 Step 5 Preview 37.8 Apex-Agents
28 Aug 2026 Tencent Hy4 preview 37.1 APEX-Agents (pass@1)
12 Aug 2026 Grok 4.6 57.5 APEX-Agents
11 Aug 2026 Solar Pro 4 18.7 APEX-Agents* (in-house)
22 Jul 2026 Solar Open 2 16.6 APEX-Agents
16 Jul 2026 Kimi K3 41 APEX-Agents
2 Jul 2026 Tencent Hy3 25.6 Apex-Agent (pass@1)
1 Jun 2026 MiniMax-M3 27.7 Apex-Agents
20 Apr 2026 Kimi K2.6 27.9 APEX-Agents
19 Feb 2026 Gemini 3.1 Pro 33.5 APEX-Agents

5 system cards report it, from 5 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
12 Aug 2026 Grok 4.6 model card 57.5 APEX-Agents, Pass@1, high, Mercor harness (3.3) Capabilities
27 Jul 2026 Kimi K3 technical report 41 APEX-Agents, cited from Artificial Analysis (Table 2) Capabilities
22 Jul 2026 Solar Open 2 technical report 16.6 APEX-Agents (Table 3) Capabilities
23 Jun 2026 Seed2.1 Pro model card 33.8 APEX Agents (Table 1) Capabilities
19 Feb 2026 Gemini 3.1 Pro model card 33.5 APEX-Agents (results table, p.4) Capabilities

Versions

Replaced by APEX-Agents 1.1

More in Economic work

All 91 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). Qwen3.8-Omni-Flash 75.3%
  • 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%