OSWorld
OSWorld is a agentic benchmark tracked by BenchGecko across every frontier and open-weight model.
OSWorld is a agentic benchmark tracked by BenchGecko across every frontier and open-weight model.
Basic
OSWorld is a agentic benchmark tracked by BenchGecko across every frontier and open-weight model.
Deep
OSWorld is a agentic benchmark tracked by BenchGecko across every frontier and open-weight model. BenchGecko tracks OSWorld scores for every eligible frontier and open-weight model. The metric is % with a maximum of 100. Category · agentic. Source · public leaderboard. See the live leaderboard for the current top 10.
Expert
OSWorld is a agentic benchmark tracked by BenchGecko across every frontier and open-weight model. Technical details: the benchmark uses % scored against 100 as max. Primary category is agentic. Source ingestion runs through the public leaderboard with the update cadence documented on the methodology page. Correlations with downstream capability have been studied in the public literature and the benchmark's authors' release notes.
Depending on why you're here
- ·OSWorld measures agentic capability with % scoring
- ·Source: public leaderboard · tracked on BenchGecko's /benchmark page
- ·Used to compare frontier models on agentic-specific tasks
- ·Pick models with high OSWorld if your workload matches agentic
- ·Benchmark scores correlate with real-world quality only for matched task types
- ·Check the live leaderboard before locking in a model · rankings shift weekly
- ·OSWorld is one of the citations labs use in launch announcements
- ·Saturation at the top of the leaderboard signals the benchmark is aging
- ·Watch for new benchmarks when all frontier models cluster within 2 points
- ·OSWorld is a test that scores how good an AI is at agentic
- ·Higher score = better model on that specific kind of task
- ·Not every score matters for every use · match the test to your goal
OSWorld is tracked live on BenchGecko · see every score across every model.