SWE-bench Verified
SWE-bench Verified evaluates AI agents on real-world programming tasks from open-source repositories sourced from GitHub, focusing on code generation and bug fixing capabilities. The benchmark involves giving agents a code repository and issue description, and challenging them to generate a patch that resolves the problem described by the issue. SWE-bench Verified is a subset of the original test set from SWE-bench, consisting of 500 samples verified by software engineers.
Paper:
SWE-bench: Can Language Models Resolve Real-World GitHub Issues? (Jimenez et al., 2023)
OpenAI Blog:
SWE-bench Verified: Introducing SWE-bench Verified
Key Features of SWE-bench Verified
Real-World Tasks
All tasks are sourced from actual GitHub issues, representing real software engineering problems.
Expert Validation
Every task has been reviewed and validated by software engineers to be non-problematic.
Diverse Tasks
Tasks originate from PRs of 12 open-source Python repositories covering various domains.
SWE-Bench Verified Leaderboard
| Rank | Agent | Models |
Verified
Verified Results
Results have been reproduced by the HAL team |
Accuracy
Accuracy
Confidence intervals show the min-max values across runs for those agents where multiple runs are available |
Cost (USD)
Total Cost
Total API cost for running the agent on all tasks. Confidence intervals show the min-max values across runs for those agents where multiple runs are available |
Runs
Number of Runs
The number of runs for this agent submitted to the leaderboard. To submit multiple evaluations, rerun the same agent and set the same agent name |
Traces |
|---|---|---|---|---|---|---|---|
| 1 | claude-3-5-sonnet-20241022 | ✓ | 38.00% | $67.09 | 1 | Download | |
| 2 | gpt-4o-2024-08-06 | ✓ | 29.80% | $79.84 | 1 | Download | |
| 3 | o1-mini-2024-09-12 | ✓ | 27.20% | $366.81 | 1 | Download |
Accuracy vs. Cost Frontier for SWE-Bench Verified
This plot shows the relationship between an agent's performance and its token cost. The Pareto frontier (dashed line) represents the current state-of-the-art trade-off. The error bars indicate min-max values across runs.
Heatmap for SWE-Bench Verified
The heatmap visualizes success rates across tasks and agents. Colorscale shows the fraction of times a task was solved across reruns of the same agent. The "any agent" performance indicates the level of saturation of the benchmark and gives a sense of overall progress.
Failure Analysis (Experimental)
Select an agent to see a detailed breakdown of failure categories and their descriptions. This analysis helps understand common failure patterns and areas for improvement. Failure reports are usually available for the top 2 agents.
Failure Categories
Distribution of Failures
Token Pricing Configuration
Adjust token prices to see how they affect the total cost calculations in the leaderboard and plots.
Additional Resources
Getting Started
Want to evaluate your agent on SWE-bench? Follow our comprehensive guide to get started:
View DocumentationTask Details
Browse the complete list of SWE-bench tasks, including problem descriptions and test cases:
View Tasks