Computer-Using Agent (OpenAI) 23 Jan 2025
OpenAI Computer-Using Agent (CUA)
OpenAI Computer-Using Agent (CUA) ranks first on 2 of the 9 benchmarks it has scores on.
13
Benchmarks with a score
2
Ranked first, out of five or more models
3
In the top three
OpenAI Jan 2025
OpenAI Computer-Using Agent (CUA)
Closed weights Flagship
No Epoch Capabilities Index score yet
- WebArena 58.1% #6 /8
- WebVoyager 87% #4 /11
- StrongREJECT 66% #17 /22
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.
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| WebVoyager | 87% | #4 of 11 | Model card | |
| OSWorld | 38.1% | #5 of 12 | Model card | |
| WebArena | 58.1% | #6 of 8 | Model card |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Human Sourced Jailbreaks | 100% | #1 of 8 | Model card | |
| Production Jailbreaks | 100% | #1 of 5 | Model card | |
| Standard Refusal Evaluation | 100% | #2 of 8 | Model card | |
| Challenging Refusal Evaluation | 92% | #4 of 9 | Model card | |
| Proactive refusals of high-risk tasks | 94% | #1 of 2 | Model card | |
| User Confirmations During Computer Use | 92% | #1 of 2 | Model card | |
| StrongREJECT | 66% | #17 of 22 | Model card | |
| Jailbreak Augmented Examples | 100% | #4 of 5 | Model card | |
| Agentic harmful task refusals | 97% | #1 of 1 | Model card | |
| Operator prompt injection scenarios | 23% | #1 of 1 | Model card |
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. 12 of them aren't in the launch post.
Operator System Card 23 Jan 2025
Capabilities
- OSWorld 38.1