BenchmarkActivecategory · codingfamily · swe-bench · variant full

SWE-bench (full test split)

swebench.com

resolve real GitHub issues (2,294 instances)

data quality57

Updated 6 h ago · first seen 11 Sept 2026

Metric: resolved · % ↑
Current results: 21
Models: 14
Current leader: claude-35-sonnet 18.1%

Filters

Comparability group

Trust level

Views

Frontier over time · resolved · board=Test · system=SWE-agent

18.1%claude-35-sonnet Anthropic Official board20 Jun 2024
12.5%GPT-4 (1106) OpenAI Official board2 Apr 2024

Includes closed rows (history). A point is emitted whenever a result beats every earlier result of the same group, ordered by evaluated_at when the source publishes it, else observed_at.

Leaderboard 4 models · trust official-benchmark

Select models with +, then Compare.

Leaderboard
#	Model	Score	Trust	Configuration	vs leader	Evaluated	Source	Actions
1	claude-35-sonnetClosedAnthropic · Claude 35	18.1%	Official board	group defaults	leader	20 Jun 2024	swebench.comT2	History
2	GPT-4 (1106)OpenAI · GPT 4	12.5%	Official board	group defaults	Partially comparable-5.66 pt	2 Apr 2024	swebench.comT2	History
3	gpt-4oClosedOpenAI · GPT 4	12.0%	Official board	group defaults	Partially comparable-6.14 pt	28 Jul 2024	swebench.comT2	History
4	claude-3-opusClosedAnthropic · Claude 3	10.5%	Official board	group defaults	Partially comparable-7.62 pt	2 Apr 2024	swebench.comT2	History

4 results

One row per canonical model — its best current row inside this comparability group (effort variants are folded into the model). Bars are relative to the page's best score. “vs leader” reads comparability: partially comparable = same task, conditions differ (reasoning effort, temperature, judge). Rules →

Definition

Task: Source:AI Atlas curated registry (YAML, versioned in git, every entry carries its source URL)T2observed 16 h agomedium
Metric: Source:AI Atlas curated registry (YAML, versioned in git, every entry carries its source URL)T2observed 16 h agomedium
Metric range: Source:AI Atlas curated registry (YAML, versioned in git, every entry carries its source URL)T2observed 6 h agomedium
Variant: Source:AI Atlas curated registry (YAML, versioned in git, every entry carries its source URL)T2observed 6 h agomedium
Harness: Source:AI Atlas curated registry (YAML, versioned in git, every entry carries its source URL)T2observed 6 h agomedium
Comparability note: Source:AI Atlas curated registry (YAML, versioned in git, every entry carries its source URL)T2observed 6 h agomedium
Paper: Source:AI Atlas curated registry (YAML, versioned in git, every entry carries its source URL)T2observed 16 h agomedium
Website: swebench.com
Source:AI Atlas curated registry (YAML, versioned in git, every entry carries its source URL)T2observed 16 h agomedium

Active group

resolved · board=Test · system=SWE-agent

4 models · 4 current rows

Comparability

Comparable same task and conditions · Partially comparable same task, conditions differ (effort, temperature, judge) · Not comparable different variant or metric

Also known as

SWE-bench (full test split), SWE-bench Full, SWE-bench Test

slug swe-bench-full · bench_01M293SPG8K6FQ2ZZD4YS1NK9R

Rows are grouped by comparability group = canonical metric × config_key (hash of the task-defining configuration keys: variant, evaluator, harness, shots, pass regime…). A leaderboard shows one row per canonical model: its best current row inside the group. Reasoning effort, temperature or judge differences keep rows in the same group but mark them partially comparable. delta_rank compares with the ranking built from the closed (previous) rows of the same group.