SWE-bench (Original Full & Lite): what it measures

One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.
		Evidence	What the result shows
Claude 3 Opus →Anthropic	3.79% 1 report	SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval	Leads this displayed comparison. Source notes Source cell: 3.79
Claude 2 →Anthropic	1.97% 1 report	SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval	#2 of 6 in this displayed comparison; 1.82 pp below the leader. Source notes Source cell: 1.97
GPT-4-turbo →OpenAI	1.31% 1 report	SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval	#3 of 6 in this displayed comparison; 2.48 pp below the leader. Source notes Source cell: 1.31
SWE-Llama 13b →Meta	0.7% 1 report	SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval	#4 of 6 in this displayed comparison; 3.09 pp below the leader. Source notes Source cell: 0.70
SWE-Llama 7b →Meta	0.7% 1 report	SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval	#4 of 6 in this displayed comparison; 3.09 pp below the leader. Source notes Source cell: 0.70
ChatGPT-3.5 →	0.17% 1 report	SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval	#6 of 6 in this displayed comparison; 3.62 pp below the leader. Source notes Source cell: 0.17
Claude 3 Opus →Anthropic	4.33% 1 report	SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval	Leads this displayed comparison. Source notes Source cell: 4.33
Claude 2 →Anthropic	3% 1 report	SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval	#2 of 6 in this displayed comparison; 1.33 pp below the leader. Source notes Source cell: 3.00
GPT-4-turbo →OpenAI	2.67% 1 report	SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval	#3 of 6 in this displayed comparison; 1.66 pp below the leader. Source notes Source cell: 2.67
SWE-Llama 7b →Meta	1.33% 1 report	SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval	#4 of 6 in this displayed comparison; 3 pp below the leader. Source notes Source cell: 1.33
SWE-Llama 13b →Meta	1% 1 report	SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval	#5 of 6 in this displayed comparison; 3.33 pp below the leader. Source notes Source cell: 1.00
ChatGPT-3.5 →	0.33% 1 report	SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval	#6 of 6 in this displayed comparison; 4 pp below the leader. Source notes Source cell: 0.33

One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.

Evidence

What the result shows

Claude 3 Opus →Anthropic

3.79%

1 report

SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval

Leads this displayed comparison.

Source notes

Source cell: 3.79

Claude 2 →Anthropic

1.97%

1 report

SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval

#2 of 6 in this displayed comparison; 1.82 pp below the leader.

Source notes

Source cell: 1.97

GPT-4-turbo →OpenAI

1.31%

1 report

SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval

#3 of 6 in this displayed comparison; 2.48 pp below the leader.

Source notes

Source cell: 1.31

SWE-Llama 13b →Meta

0.7%

1 report

SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval

#4 of 6 in this displayed comparison; 3.09 pp below the leader.

Source notes

Source cell: 0.70

SWE-Llama 7b →Meta

0.7%

1 report

SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval

#4 of 6 in this displayed comparison; 3.09 pp below the leader.

Source notes

Source cell: 0.70

ChatGPT-3.5 →

0.17%

1 report

SWE-bench (Original Full & Lite) · Full · resolved rate[1]SWE-bench Full · BM25 retrieval

#6 of 6 in this displayed comparison; 3.62 pp below the leader.

Source notes

Source cell: 0.17

Claude 3 Opus →Anthropic

4.33%

1 report

SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval

Leads this displayed comparison.

Source notes

Source cell: 4.33

Claude 2 →Anthropic

1 report

SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval

#2 of 6 in this displayed comparison; 1.33 pp below the leader.

Source notes

Source cell: 3.00

GPT-4-turbo →OpenAI

2.67%

1 report

SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval

#3 of 6 in this displayed comparison; 1.66 pp below the leader.

Source notes

Source cell: 2.67

SWE-Llama 7b →Meta

1.33%

1 report

SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval

#4 of 6 in this displayed comparison; 3 pp below the leader.

Source notes

Source cell: 1.33

SWE-Llama 13b →Meta

1 report

SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval

#5 of 6 in this displayed comparison; 3.33 pp below the leader.

Source notes

Source cell: 1.00

ChatGPT-3.5 →

0.33%

1 report

SWE-bench (Original Full & Lite) · Lite · resolved rate[1]SWE-bench Lite · BM25 retrieval

#6 of 6 in this displayed comparison; 4 pp below the leader.

Source notes

Source cell: 0.33

Sources

SWE-bench: Can Language Models Resolve Real-World GitHub Issues? — published result table ↗

Jimenez, Carlos E., Yang, John, Wettig, Alexander, Yao, Shunyu, Pei, Kexin, Press, Ofir et al. · 2023

All 7 authors

Jimenez, Carlos E., Yang, John, Wettig, Alexander, Yao, Shunyu, Pei, Kexin, Press, Ofir, Narasimhan, Karthik

Dates & evaluation details

SWE-bench (Original Full & Lite) · resolved rate · SWE-bench Full · BM25 retrieval · 2026-10-26

Published model comparison, 2023. SWE-bench Full · BM25 retrieval. Test-verified issue resolution with the original BM25 retrieval scaffold. Full and Lite are different task sets and remain separate. Table 5; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

SWE-bench (Original Full & Lite) · resolved rate · SWE-bench Lite · BM25 retrieval · 2026-10-26

Published model comparison, 2023. SWE-bench Lite · BM25 retrieval. Test-verified issue resolution with the original BM25 retrieval scaffold. Full and Lite are different task sets and remain separate. Table 5; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.

TRACE THE SCORE · Per-model notes use published breakdowns where available.

What it measures

Full 2,294-issue benchmark and 300-issue SWE-bench Lite subset.

Reading the measurement

Look beyond a plausible code sample to the executable artifact. Tests, repository context, language coverage and time limits shape the result. Separate first-attempt correctness from success after debugging, retries or agent assistance.

Published evaluation

Test-verified issue resolution with the original BM25 retrieval scaffold. Full and Lite are different task sets and remain separate.

Where this fits in an evaluation suite

The agent wrapper is part of the system. Pin tools, dependencies, timeouts, tests, and the repository revision. Compare contest coding separately from repository maintenance.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Test-verified issue resolution with the original BM25 retrieval scaffold. Full and Lite are different task sets and remain separate.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

SWE-bench: Can Language Models Resolve Real-World GitHub Issues? — published result table ↗
Jimenez, Carlos E., Yang, John, Wettig, Alexander, Yao, Shunyu, Pei, Kexin, Press, Ofir et al. · 2023
All 7 authors
Jimenez, Carlos E., Yang, John, Wettig, Alexander, Yao, Shunyu, Pei, Kexin, Press, Ofir, Narasimhan, Karthik

Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.

Sources

What it measures

Reading the measurement

Published evaluation

Where this fits in an evaluation suite

A reproducible comparison

Continue exploring

AA-3D

Terminal-Bench 4.0

FrontierCode 1.1

Terminal-Bench 2.1