TheAgentCompany
Office tasks in a simulated software company.
As of 19 Sept 2026, DeepSeek V3.2 leads TheAgentCompany on BenchLeader with 42.9%, ahead of Gemini 2.5 Flash 09 2025 at 41.1%, across 14 model configurations with a published result.
- Published by
- TheAgentCompanydata via Epoch AI Benchmarking Hub
- Category
- Agents & tools
- Index weight
- Reference only
- Models
- 14
- Data as of
- 19 Sept 2026
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
What the test looks like
An agent works in a simulated company with chat, code hosting and office tools, completing tasks a new hire would.
How it is scored
Percent of tasks resolved, as published by TheAgentCompany.
What to keep in mind
Harness-dependent; the environment is simulated.
- 1DeepSeek V3.242.9%
- 2Gemini 2.5 Flash 09 202541.1%
- 3Claude Sonnet 433.1%
- 4Claude 3.7 Sonnet30.9%
- 5Gemini 2.5 Pro30.3%
- 6Claude 3.5 Sonnet24.0%
- 7Gemini 2.0 Flash11.4%
- 8GPT-4o8.6%
- 9Llama 3.1 405B7.4%
- 10Llama 3.1 70B6.9%
- 11Qwen2.5 72B5.7%
- 12Gemini 1.5 Pro 0023.4%
- 13Nova Pro1.7%
- 14Qwen2-72B1.1%
14 of 14
| # | ||||
|---|---|---|---|---|
| 1 |  DeepSeek V3.2DeepSeekopen | 42.9% | 56.0 | 2025-09-29 |
| 2 |  Gemini 2.5 Flash 09 2025Google | 41.1% | 52.1 | 2025-09-25 |
| 3 |  Claude Sonnet 4Anthropic | 33.1% | 49.7 | 2025-05-22 |
| 4 | Claude 3.7 SonnetAnthropic | 30.9% | 50.2 | 2025-02-24 |
| 5 | Gemini 2.5 ProGoogle | 30.3% | 54.5 | 2025-05-06 |
| 6 | Claude 3.5 SonnetAnthropic | 24.0% | 46.6 | 2024-10-22 |
| 7 | Gemini 2.0 FlashGoogle | 11.4% | 43.7 | 2025-02-05 |
| 8 |  GPT-4oOpenAI | 8.6% | 44.3 | 2024-11-20 |
| 9 |  Llama 3.1 405BMetaopen | 7.4% | 43.5 | 2024-07-23 |
| 10 | Llama 3.1 70BMetaopen | 6.9% | 41.3 | 2024-07-23 |
| 11 |  Qwen2.5 72BAlibabaopen | 5.7% | 42.7 | 2024-09-19 |
| 12 | Gemini 1.5 Pro 002Google | 3.4% | 44.9 | 2024-09-24 |
| 13 |  Nova ProAmazon | 1.7% | 41.4 | 2024-12-03 |
| 14 | Qwen2-72BAlibabaopen ↗ | 1.1% | 41.3 | 2024-06-07 |
Cite as: BenchLeader, “TheAgentCompany leaderboard”, https://www.benchleader.com/benchmarks/the_agent_company, data as of 19 Sept 2026.
TheAgentCompany: questions
- What does TheAgentCompany measure?
- An agent works in a simulated company with chat, code hosting and office tools, completing tasks a new hire would. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads TheAgentCompany?
- DeepSeek V3.2 leads TheAgentCompany with 42.9% as of 19 Sept 2026, ahead of Gemini 2.5 Flash 09 2025 at 41.1%.
- How many models have TheAgentCompany results?
- 14 model configurations have a TheAgentCompany result on BenchLeader, all taken from TheAgentCompany via Epoch AI Benchmarking Hub.
- Who runs TheAgentCompany and how often is it updated?
- TheAgentCompany is published by TheAgentCompany. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does TheAgentCompany count toward the BenchLeader Index?
- No. TheAgentCompany is shown for reference but left out of the composite index.