SWE-bench

SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

↓ SWE-bench Verified

2,294 task instances built from real GitHub issues and their resolving pull requests across 12 popular Python repositories. Given the issue text and a repository snapshot, an agent must produce a patch; it is scored by running the PR's FAIL_TO_PASS and PASS_TO_PASS unit tests. The original agentic-coding benchmark and the template for the Verified, Lite, Multimodal, Multilingual and Pro descendants.

52.62%
Sonar Foundation Agent + Claude 4.5 Opus
official
2023-10-10 → 2025-12-19: 1.96% → 52.62%
Released
2023-10
Maintainer
Princeton NLP / SWE-bench team (Jimenez et al.)
Status
saturated
Contamination
high
Metric
% resolved (percent, ↑)
Tasks
2,294
Domains
software-engineering code tool-use
human
no measured baseline

Notes. The full 2,294-instance set contains under-specified and untestable tasks (the reason OpenAI and the maintainers curated SWE-bench Verified). Almost all frontier evaluation moved to Verified after August 2024; the Full leaderboard still accepts submissions but receives few, so treat it as historical.

Full ledger

SystemDeveloperScoreDateSourceConditions
Sonar Foundation Agent + Claude 4.5 OpusAnthropic52.62%officialsplit: full scaffold: Sonar Foundation Agent
Top of the Full leaderboard; submission not run or checked by the SWE-bench team.
SWE-agent 1.0 (Claude 3.7 Sonnet)Anthropic33.83%officialsplit: full scaffold: SWE-agent 1.0
OpenHands + CodeAct v2.1 (claude-3-5-sonnet-20241022)Anthropic29.38%officialsplit: full scaffold: OpenHands CodeAct v2.1
SWE-agent + GPT 4 (1106)OpenAI12.47%officialsplit: full scaffold: SWE-agent
RAG + Claude 2Anthropic1.96%officialsplit: full scaffold: RAG baseline
Paper baseline (BM25 retrieval, no agent), checked by the SWE-bench team.