Inference
EN
SWE-bench team

SWE-bench Multilingual · SWE-bench/experiments

Resolved-issue rate on issues across nine programming languages, from the submission records behind the leaderboard.

ModelsModels with a result.14
Top resultBest result on this benchmark.72.7%Gemini 3 Flash
Top-3 spreadPoints from first to third.2.0 pts
UpdatedDate the source changed the results.2 Sep 2026

Result and cost

14 of 14 results

Results

14 results
#ModelResultThe score from the source.CostUS dollars to run one task.
01Gemini 3 Flashmini-SWE-agent72.7%$0.352
02Claude Opus 4.6mini-SWE-agent72.0%$0.663
03Claude Opus 4.5mini-SWE-agent70.7%$0.834
04GLM-5mini-SWE-agent69.7%$0.642
05Gemini 3 Promini-SWE-agent68.7%$1.02
06MiniMax 2.5mini-SWE-agent68.3%$0.101
07Kimi K2.5mini-SWE-agent67.3%$0.693
08Gemini 3.5 Flashmedium effort · mini-SWE-agent67.0%$0.627
08Claude Sonnet 4.5mini-SWE-agent67.0%$0.67
10GPT-5.2high effort · mini-SWE-agent66.7%$0.536
11GPT-5.2-Codexmini-SWE-agent66.3%$0.662
12Claude Haiku 4.5mini-SWE-agent64.7%$0.379
13DeepSeek V3.2mini-SWE-agent59.0%$0.384
14GPT-5 minimini-SWE-agent39.7%$0.052