Developer

Refactoring assistant

Refactor code safely while preserving behavior and improving clarity.

task.code_refactortask.code_review

Evidence quality is currently limited for this use case. Rankings below are useful for exploration, not a strong winner claim.

Provisional leader

claude-sonnet-4

Current leader based on limited benchmark evidence. Treat this ranking as directional until coverage improves.

31.2%

Best benchmark score

43.8%

Confidence

All ranked models — top 3

🥇

claude-sonnet-4

31.2%

🥈

gpt-5-2025-08-07

27.3%

🥉

Kimi K2 Thinking

24.5%

Ranked Models

Evidence Quality

82%

Evidence Points

Top Signal

SWE-bench Verified Leaderboard: swe_verified_resolved_pct

All Ranked Models

30 of 30 models

Rank	Model	Score	Confidence	Price / 1M	Evidence sources
🥇	claude-sonnet-4 Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Galileo Agent Leaderboard v2 Avg AC	31.2%	44%	$6.00	SWE-bench Verified LeaderboardGalileo Agent Leaderboard v2
🥈	gpt-5-2025-08-07 Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Aider Polyglot Leaderboard percent_correct_pct	27.3%	33%	—	SWE-bench Verified LeaderboardAider Polyglot Leaderboard
#4	Kimi K2 Thinking Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Sonar Java Quality Leaderboard functional_skill_pct	24.5%	35%	$1.07	SWE-bench Verified LeaderboardSonar Java Quality Leaderboard
#5	gpt-4o-2024-05-13 Strong on RepoQA Official Results overall_average_pass_at_1_pct and SWE-bench Verified Leaderboard swe_verified_resolved_pct	23.6%	32%	—	RepoQA Official ResultsSWE-bench Verified Leaderboard
#7	gemini-3-pro-preview Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Vals SWE-bench overall_accuracy_pct	20.9%	25%	$4.50	SWE-bench Verified LeaderboardVals SWE-bench
#8	gemini-2.5-pro Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Galileo Agent Leaderboard v2 Avg AC	20.9%	30%	$3.44	SWE-bench Verified LeaderboardGalileo Agent Leaderboard v2
#9	gpt-5.2-2025-12-11 Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Vals SWE-bench overall_accuracy_pct	20.8%	24%	—	SWE-bench Verified LeaderboardVals SWE-bench
#10	o3-20250416 Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Aider Polyglot Leaderboard percent_correct_pct	19.9%	25%	$3.50	SWE-bench Verified LeaderboardAider Polyglot Leaderboard
#11	gpt-5-mini-2025-08-07 Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Vals LiveCodeBench overall_accuracy_pct	18.8%	28%	—	SWE-bench Verified LeaderboardVals LiveCodeBench
#12	glm-4.7 Strong on Sonar Java Quality Leaderboard functional_skill_pct and Vals LiveCodeBench overall_accuracy_pct	18.4%	27%	—	Sonar Java Quality LeaderboardVals LiveCodeBench
#13	deepseek-r1 Strong on Sonar Java Quality Leaderboard functional_skill_pct and Aider Polyglot Leaderboard percent_correct_pct	18.1%	26%	$0.27	Sonar Java Quality LeaderboardAider Polyglot Leaderboard
#15	claude-opus-4-5-20251101 Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Vals Terminal-Bench 2 overall_accuracy_pct	18.0%	21%	—	SWE-bench Verified LeaderboardVals Terminal-Bench 2
#16	gpt-4.1-20250414 Strong on Galileo Agent Leaderboard v2 Avg AC and SWE-bench Verified Leaderboard swe_verified_resolved_pct	17.8%	28%	—	Galileo Agent Leaderboard v2SWE-bench Verified Leaderboard
#17	minimax-m2.1 Strong on Sonar Java Quality Leaderboard functional_skill_pct and Vals SWE-bench overall_accuracy_pct	17.7%	27%	$0.53	Sonar Java Quality LeaderboardVals SWE-bench
#18	gpt-4o-20241120 Strong on Aider Code Editing Leaderboard percent_correct_pct and BigCodeBench Official bigcodebench_complete_pct	17.4%	31%	—	Aider Code Editing LeaderboardBigCodeBench Official
#19	gemini-3.1-pro-preview Strong on Vals SWE-bench overall_accuracy_pct and Vals Terminal-Bench 2 overall_accuracy_pct	16.9%	18%	$4.50	Vals SWE-benchVals Terminal-Bench 2
#20	claude-opus-4 Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Aider Polyglot Leaderboard percent_correct_pct	16.5%	20%	$10.00	SWE-bench Verified LeaderboardAider Polyglot Leaderboard
#21	o4-mini Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Aider Polyglot Leaderboard percent_correct_pct	15.9%	24%	$1.93	SWE-bench Verified LeaderboardAider Polyglot Leaderboard
#24	qwen-2.5-72b-instruct Strong on Galileo Agent Leaderboard v2 Avg AC and Aider Code Editing Leaderboard percent_correct_pct	15.3%	21%	—	Galileo Agent Leaderboard v2Aider Code Editing Leaderboard
#25	gpt-4.1 Strong on SWE-bench Verified Leaderboard swe_verified_resolved_pct and Aider Polyglot Leaderboard percent_correct_pct	15.2%	18%	$3.50	SWE-bench Verified LeaderboardAider Polyglot Leaderboard
#26	Grok-4-0709 Strong on Vals LiveCodeBench overall_accuracy_pct and Galileo Agent Leaderboard v2 Avg AC	14.9%	23%	—	Vals LiveCodeBenchGalileo Agent Leaderboard v2
#27	gpt-4o Strong on Sonar Java Quality Leaderboard functional_skill_pct and SWE-bench Verified Leaderboard swe_verified_resolved_pct	14.6%	32%	$0.26	Sonar Java Quality LeaderboardSWE-bench Verified Leaderboard
#30	gemini-3-flash-preview Strong on Vals SWE-bench overall_accuracy_pct and Vals LiveCodeBench overall_accuracy_pct	14.2%	17%	$1.13	Vals SWE-benchVals LiveCodeBench
#32	gpt-5.4-2026-03-05 Strong on Vals SWE-bench overall_accuracy_pct and Vals LiveCodeBench overall_accuracy_pct	14.1%	16%	—	Vals SWE-benchVals LiveCodeBench
#33	claude-sonnet-4.6 Strong on Vals SWE-bench overall_accuracy_pct and Vals LiveCodeBench overall_accuracy_pct	13.8%	16%	$6.00	Vals SWE-benchVals LiveCodeBench
#34	claude-opus-4-6-thinking Strong on Vals SWE-bench overall_accuracy_pct and Vals LiveCodeBench overall_accuracy_pct	13.5%	15%	—	Vals SWE-benchVals LiveCodeBench
#37	GLM-5 Strong on Sonar Java Quality Leaderboard functional_skill_pct and Sonar Java Quality Leaderboard issue_density_error_per_kloc	13.3%	15%	—	Sonar Java Quality LeaderboardSonar Java Quality Leaderboard
#38	gpt-5.1-2025-11-13 Strong on Vals LiveCodeBench overall_accuracy_pct and Vals SWE-bench overall_accuracy_pct	13.0%	16%	—	Vals LiveCodeBenchVals SWE-bench
#39	claude-opus-4-5-20251101-thinking Strong on Vals SWE-bench overall_accuracy_pct and Vals LiveCodeBench overall_accuracy_pct	13.0%	15%	—	Vals SWE-benchVals LiveCodeBench
#43	gpt-4o-2024-08-06 Strong on Aider Code Editing Leaderboard percent_correct_pct and Aider Code Editing Leaderboard correct_edit_format_pct	11.9%	20%	—	Aider Code Editing LeaderboardAider Code Editing Leaderboard

Compare Models

Select two different models above to compare their evidence side by side.

▶Ranking diagnostics & missing models

Source lift

Ranked

Sources

Quality

Low

Vals LiveCodeBench

39 rows · 1.1% avg lift

Vals Legal Bench

39 rows · 0.2% avg lift

Vals Tax Eval v2

36 rows · 0.2% avg lift

Vals MedQA

34 rows · 0.2% avg lift

Missing frontier models

No obvious gaps right now.

▶Taxonomy & task details

Core tasks

task.code_refactortask.code_review

Required modes

none

Domains

domain.software_engineering

Related in Developer

Codebase onboarding brief

Summarize a repository's architecture, modules, and conventions.

Autonomous Coding Agent

End-to-end autonomous software engineering: reading issues, writing code, running tests, submitting PRs.

Code generation

Generate correct, secure code from requirements.

IDE code completion

Fast local-context code completion and small snippet generation.