- Humanity's Last Exam (Full Set), browser + computer + terminal 41.6
- FrontierMath Tier 1-3, with tools (text) 27.4
- DSBench: Data Analysis 89.9 (Data Modeling 85.5) 89.9
- SpreadsheetBench, OSX LibreOffice, with .xlsx access 45.5 (without 35.3) 45.5
- Investment Banking Modeling Tasks, mean accuracy (Oracle@64 71.3) 41
- BrowseComp (Agentic Browsing) 68.9
- WebArena (Agentic Browser Use) 65.4
ChatGPT agent
ChatGPT agent ranks first on 4 of the 13 benchmarks with five or more models scored.
23
Benchmarks with a score
4
Ranked first, out of five or more models
5
In the top three
OpenAI Jul 2025
ChatGPT agent
Closed weights Flagship
155.9
#54 of 404
Self-Reported Index, from the lab's own numbers. No Epoch Capabilities Index score yet.
- Remote Labor Index 1.25% #17 /19
- WebArena 65.4% #3 /8
- CocoaBench 26.8% #4 /6
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SimpleQA | 91.4% | #5 of 60 | Lab model card | |
| SimpleQA Hallucination Rate | 7.90% | #1 of 12 | Lab model card | |
| PersonQA | 94.3% | #1 of 9 | Lab model card | |
| Humanity's Last Exam | 41.6% | #62 of 154 | Lab model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| FrontierMath | 27.4% | #7 of 12 | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| DSBench | 89.9% | #1 of 1 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| WebArena | 65.4% | #3 of 8 | Lab model card | |
| BrowseComp | 68.9% | #47 of 81 | Lab model card | |
| CocoaBench | 26.8% | #4 of 6 | Official leaderboard | |
| AgentIF-OneDay | 62.6% | #1 of 1 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MBABench | 31.2 | #1 of 6 | Official leaderboard | |
| SpreadsheetBench | 45.5% | #5 of 6 | Lab model card | |
| Remote Labor Index | 1.25% | #17 of 19 | Official leaderboard | |
| Investment Banking Modeling Tasks (OpenAI) | 41% | #5 of 5 | Lab model card |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| First-Person Fairness evaluation | 0.40% | #1 of 10 | Lab model card | |
| Filtered, adversarial sample of production prompts | 96.9% | #1 of 3 | Lab model card | |
| Challenging prompts from red teamers with biosafety-relevant PhDs | 87.9% | #3 of 3 | Lab model card | |
| Proactive refusals of high-risk tasks | 89% | #2 of 2 | Lab model card | |
| User Confirmations During Computer Use | 91% | #2 of 2 | Lab model card | |
| Active data exfiltration, visual browser | 67% | #1 of 1 | Lab model card | |
| In-context data exfiltration, visual browser | 78% | #1 of 1 | Lab model card | |
| Irrelevant instructions, text-based web browser | 99.5% | #1 of 1 | Lab model card | |
| Irrelevant instructions, visual browser | 95% | #1 of 1 | Lab model card |
Its lab's numbers put it at 155.9, number 54 of 404
The Self-Reported Index fits one score to the 5 capability results OpenAI reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 150.0 to 161.7. How we compute it.
- PersonQA 94.3
- SimpleQA 91.4
- BrowseComp 68.9
- Humanity's Last Exam 41.6
- FrontierMath 27.4
Its lab's number next to a leaderboard
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| SpreadsheetBench | 45.5% SpreadsheetBench, OSX LibreOffice, with .xlsx access 45.5 (without 35.3) | 35.3% Leaderboard | +10 pts |
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.
Its system card
Every result in the longer documents the lab published with it. 32 of them aren't in the launch post.
ChatGPT Agent System Card Jul 17, 2025
Hallucination
- SimpleQA 91.4
- SimpleQA Hallucination Rate 7.9
- PersonQA 94.3
Safety
- Standard Refusal Evaluation
- OpenAI Production Benchmarks and safe-completion evals
- Image input safety evaluations
- User Confirmations During Computer Use 91
- Agentic harmful task refusals
- Proactive refusals of high-risk tasks 89
- Challenging prompts from red teamers with biosafety-relevant PhDs 87.9
- Filtered, adversarial sample of production prompts 96.9