Home › Benchmarks › SWE-bench Verified
SWE-bench Verified
Benchmark
Capability
Real GitHub issues, real repos — the coding-agent standard.
You're looking at a benchmark record: a test the industry uses to measure chips or models — and what it actually tells you.
SWE-bench gives a model an actual repository and issue; success means a patch that passes the tests. The Verified subset (human-validated, with OpenAI) became the metric coding agents live by — from single digits in 2023 to ~80% at the 2025 frontier — and the number behind every 'AI software engineer' claim.
How this connects to the rest of the stack
This record's links at a glance — click a group to jump to it.
Benchmarks
Sources
Record reviewed as of 2026-01-15 — see Methodology.