Skip to content

Best models for coding

69 models ranked on coding, plus the agentic software-engineering benchmarks behind it: real repository fixes, terminal tasks and multi-step engineering work.

Top coding model
Claude Fable 5.1
Anthropic
Coding 91.2
Best open-weights coder
Qwen3.8 Max
Alibaba
Coding 70.6
Best value (top 30)
MiMo-V2.6-Flash
Xiaomi
366 points per $ · $0.17 / 1M

Coding ranking

Top 25 by coding score. The official rank in the full table follows BenchLM's ranking method, so its order can differ.

Full table →
  1. 1Claude Fable 5.1Anthropic91.2
  2. 2Claude Opus 5Anthropic90.4
  3. 3Claude Fable 5Anthropic90.3
  4. 4Claude Opus 5.5Anthropic87.5
  5. 5Claude Sonnet 5.5Anthropic87.5
  6. 6Gemini 3.7 FlashGoogle78.6
  7. 7Gemini 3.1 ProGoogle78.1
  8. 8GPT-6 SolOpenAI77.5
  9. 9Gemini 3.8 FlashGoogle77.5
  10. 10Grok 4.6xAI77.3
  11. 11GPT-5.2-CodexOpenAI76.7
  12. 12GPT-6 AstraOpenAI76.4
  13. 13Claude Opus 4.8Anthropic75.1
  14. 14Kimi K3Moonshot AI74.6
  15. 15GPT-6.1 SolOpenAI74.1
  16. 16Gemini 4 ArgonGoogle74.0
  17. 17Qwen3.8 MaxAlibaba70.6
  18. 18Muse Wonder FullMeta69.0
  19. 19Claude Opus 4.7Anthropic69.0
  20. 20Grok 4.5xAI67.7
  21. 21GPT-6 LunaOpenAI66.3
  22. 22GPT-5.4 nanoOpenAI66.1
  23. 23Claude Opus 4.7 (Adaptive)Anthropic65.6
  24. 24Claude Sonnet 4.6Anthropic65.6
  25. 25Qwen3.7 MaxAlibaba65.5
Higher is better.

Coding vs price

Blended price per 1M tokens. The dashed line is the Pareto frontier.

Pareto frontierMost attractive quadrant65 models
020406080100$0.01$0.1$1$10$100Blended price, USD per 1M tokens (log scale) · lower is betterCoding · higher is betterClaude Fable 5.1Claude Sonnet 5.5Gemini 3.7 FlashGemini 3.8 FlashGPT-6 LunaMiMo-V2.6-FlashGLM-5.3-FlashLing 3.0 Flash

WebDev Arena (human votes)

People ask two anonymous models to build the same web app and vote for the better result.

#ModelRatingVotes
1Claude Opus 5.5 (Max)Anthropic18072,925
2GPT-6 Astra (Max)OpenAI17746,169
3Claude Sonnet 5.5 (xHigh)Anthropic17683,428
4GPT-6.1 Sol (Max)OpenAI17533,072
5Claude Fable 5.1 (Max)Anthropic17437,744
6Muse Wonder Full (Max)Meta17261,639
7Claude Sonnet 5.5 (High)Anthropic17105,158
8Claude Opus 5 (Max)Anthropic168918,555
9GPT-6 Sol (Max)OpenAI16834,608
10Gemini 4 Argon (High)Google16781,728
11Qwen3.8-Max (0902)Alibaba167414,421
12Qwen3.8-MaxAlibaba16723,428
13DeepSeek-V4.1-Flash-20261017 (Max)DeepSeek16642,008
14Muse Spark 1.3 (Max)Meta16599,407
15Claude Opus 5 (High)Anthropic165924,739

Showing the top 15 of 148.

SWE-bench Pro

A long-horizon repository benchmark built to test realistic software engineering work. Its scores need a task-quality and setup check before they support a coding-agent decision.

All 69 →
  1. 1Claude Opus 5.589.9
  2. 2Claude Sonnet 5.581.3
  3. 3Claude Fable 5.181.2
  4. 4Claude Mythos 580.3
  5. 5Claude Fable 580.0
  6. 6Claude Opus 579.2
  7. 7Claude Opus 4.869.2
  8. 8Ling 3.1 Flash68.7
  9. 9Qwen3.8 Max67.7
  10. 10Beam 1.0 501B-A23B66.1
  11. 11Hy4 preview65.7
  12. 12Ornith-1.5-397B65.1
Higher is better.

SWE-bench Verified

SWE-bench Verified tests patches for 500 human-filtered GitHub issues. A score measures how many tasks an evaluated model-and-agent setup resolved; the model name alone does not identify that setup.

All 98 →
  1. 1Claude Opus 596.0
  2. 2Claude Mythos 595.5
  3. 3Claude Fable 595.0
  4. 4Ember-192.2
  5. 5Claude Opus 4.888.6
  6. 6Claude Opus 4.7 (Adaptive)87.6
  7. 7Ornith-1.5-397B86.0
  8. 8Claude Sonnet 585.2
  9. 9GPT-5.3 Codex85.0
  10. 10Ornith-1.0-397B82.4
  11. 11Beam 1.0 501B-A23B81.6
  12. 12Claude Opus 4.580.9
Higher is better.

Terminal-Bench 2.1

Terminal-Bench 2.1 results, mostly provider-reported, stored separately from the Terminal-Bench 2.0 lane.

All 68 →
  1. 1SWE-292.8
  2. 2GPT-5.6 Sol91.9
  3. 3DeepSeek V4.1 Flash 101790.6
  4. 4MiMo-V2.6-Pro89.9
  5. 5Gemini 3.8 Flash89.4
  6. 6Muse Spark 1.388.8
  7. 7Kimi K388.3
  8. 8GLM-5.388.2
  9. 9Claude Mythos 588.0
  10. 10DeepSeek V4 Pro 081387.9
  11. 11MiMo-V2.6-Flash87.6
  12. 12GPT-5.6 Terra87.4
Higher is better.

DeepSWE

A long-horizon software engineering benchmark from Datacurve for measuring frontier coding agents on original tasks drawn from active open-source repositories.

All 50 →
  1. 1Muse Wonder Full79.0
  2. 2Muse Wonder78.8
  3. 3Gemini 4 Argon77.9
  4. 4DeepSeek V4.1 Flash 101777.0
  5. 5Muse Spark 1.375.4
  6. 6Ember-175.2
  7. 7Claude Opus 5.574.2
  8. 8GPT-6 Astra74.1
  9. 9Pareto 26.974.0
  10. 10Gemini 3.8 Flash73.8
  11. 11MiniMax M3.1 Flash73.7
  12. 12SWE-273.0
Higher is better.

FrontierSWE v2

A 34-task expansion of FrontierSWE for ultra-long-horizon engineering and research work that remains far from saturation.

All 20 →
  1. 1GPT-6 Astra65.5
  2. 2Claude Opus 5.562.3
  3. 3Claude Sonnet 5.561.9
  4. 4Claude Fable 5.156.3
  5. 5Gemini 4 Argon55.0
  6. 6Claude Opus 552.0
  7. 7Claude Fable 547.0
  8. 8Claude Haiku 5.543.8
  9. 9GPT-5.6 Sol32.2
  10. 10GLM-5.330.2
  11. 11Grok 4.729.5
  12. 12Kimi K325.9
Higher is better.

SWE Multilingual

A multilingual software-engineering benchmark for real-world code issue resolution across multiple programming languages.

All 46 →
  1. 1Claude Opus 5.593.9
  2. 2Claude Mythos 592.2
  3. 3Claude Sonnet 5.590.3
  4. 4Claude Opus 589.5
  5. 5Claude Fable 5.189.1
  6. 6Ling 3.1 Flash88.5
  7. 7Claude Opus 4.884.4
  8. 8Hy4 preview82.9
  9. 9Qwen3.8-Flash-Next81.0
  10. 10Qwen3.8-Omni-Flash80.5
  11. 11Composer 2.579.8
  12. 12Ornith-1.5-397B79.6
Higher is better.

LiveCodeBench v6

LiveCodeBench v6 is a named release slice used in provider comparison tables. Keeping it separate prevents v6 results from being mixed into older or rolling LiveCodeBench windows.

All 46 →
  1. 1Qwen3.8-Omni-Flash92.6
  2. 2Solar Open 292.4
  3. 3Qwen3.8-Flash-Next91.9
  4. 4Qwen3.7 Max91.6
  5. 5dots3-note Preview91.5
  6. 6Qwen3.8-27B90.3
  7. 7Ternary Bonsai 2 27B90.1
  8. 8Kimi K2.689.6
  9. 9Qwen3.7 Plus89.6
  10. 10Nemotron 3 Ultra89.0
  11. 11Nemotron 3.5 Super VL 120B A12B88.2
  12. 12MAI-Thinking-187.7
Higher is better.

NL2Repo

A repository-understanding benchmark that measures whether models can map natural-language requests onto the right code locations and system changes.

All 31 →
  1. 1DeepSeek V4.1 Flash 101768.2
  2. 2DeepSeek V4 Pro 081361.5
  3. 3MiniMax M3.1 Flash61.2
  4. 4Ornith-1.5-397B59.5
  5. 5Hy4 preview58.9
  6. 6GLM-5.358.0
  7. 7GLM-5.3-Flash56.3
  8. 8Qwen3.8 Max55.9
  9. 9DeepSeek V4 Flash 073154.2
  10. 10dots3-note Preview49.8
  11. 11GLM-5.248.9
  12. 12Qwen3.8-Omni-Flash48.9
Higher is better.

Results are as reported by labs and evaluators, often with different agent scaffolds, tool access and attempt counts. A head-to-head comparison of coding agents (the same model in different harnesses such as IDE or CLI agents) needs our own runs and is not published yet.