Skip to content
ModelCap

Coding ranking

Best AI models for coding

Language models measured on at least two public coding boards (Arena Coding, SWE-bench bash-only, Terminal-Bench 2.1 with Terminus 2), ordered by the coding component of the ModelCap Index. Models with a single coding result follow in their own tier, ordered by that board, so one result never outranks two.

Snapshot as of 19 November 2026

Gemini 4 Carbon from Google leads the best-for-coding list as of 19 November 2026 with Arena coding 1567 on one public coding board, holding ModelCap Index position #1; 102 models qualify.

Models on this list
102
With openly downloadable weights
41
With a listed API price
77

Best AI models for coding (0)

Every row is a current, canonical language model with a public position on the ModelCap Index; the list filters by a published fact and orders by a published figure.

No ranked model currently holds admitted results on two or more public coding boards, so the primary tier is empty at this refresh. The single-board tier below is the complete coding evidence the Index admits right now.

Measured on one coding board (top 100 of 102)

Ranked models with a single admitted coding result, ordered by that board's normalised score. One board is a narrower claim than two, so these rows sit apart from the list above and never mix into it.

Measured on one coding board: Ranked models with a single admitted coding result, ordered by that board's normalised score. One board is a narrower claim than two, so these rows sit apart from the list above and never mix into it.
#ModelCoding boardsModelCap IndexInput $/1MOutput $/1MContextWeights
1Gemini 4 CarbonGoogleArena coding 1567#187.8$2.00$10.001MNo linked weights
2Gemini 4 ArgonGoogleArena coding 1554#386.6$2.00$10.001MNo linked weights
3Gemini 4 NeonGoogleArena coding 1558#287.2$3.00$15.001MNo linked weights
4Muse WonderMetaArena coding 1549#585.6$4.00$20.001MNo linked weights
5Claude Opus 5.5AnthropicArena coding 1547 · Terminal-Bench 4.0 64.9%#485.8$4.00$20.001MNo linked weights
6Kimi K3Moonshot AIArena coding 1536#1779.0$0.192$16.001MRestricted license
7Muse Spark 1.3MetaArena coding 1535 · Terminal-Bench 4.0 14.6%#782.4$1.25$4.251MNo linked weights
8Kimi K3.1 ProMoonshot AIArena coding 1537#1180.0$0.253$8.801MRestricted license
9MiMo-V2.6-ProXiaomiArena coding 1538#2078.4$0.435$0.871MOpen weights
10Gemini 3.8 FlashGoogleArena coding 1530 · Terminal-Bench 4.0 19.1%#881.4$0.825$4.131MNo linked weights
11GLM 5.3 FlashZ.aiArena coding 1527 · Terminal-Bench 4.0 35.8%#2676.4$0.15$0.501MOpen weights
12GLM 5.3Z.aiArena coding 1523 · Terminal-Bench 4.0 41.8%#1978.5$0.07$1.401MRestricted license
13Gemini 3.1 Pro PreviewGoogleArena coding 1520#1679.2$2.00$12.001MNo linked weights
14GPT-5.5OpenAIArena coding 1518#1878.8$5.00$30.001MNo linked weights
15Claude Sonnet 5.5AnthropicArena coding 1518 · Terminal-Bench 4.0 61.8%#2477.3$2.00$10.001MNo linked weights
16Gemini 4 HydrogenGoogleArena coding 1515#980.4$1.50$7.501MNo linked weights
17MiMo-V2.6-FlashXiaomiArena coding 1511#4668.8$0.14$0.281MOpen weights
18Grok 4.20 Multi-AgentxAIArena coding 1508#2776.2$1.25$2.502MNo linked weights
19GPT-6 LunaOpenAIArena coding 1504 · Terminal-Bench 4.0 16.4%#5866.3$0.10$0.501MNo linked weights
20Beam 1.0 501B A23BReflectionArena coding 1505#6763.4$0.25$1.001MOpen weights
21Step 5 PreviewStepFunArena coding 1504#4968.4$1.00$2.701MRestricted license
22Gemma 4 31BGoogleArena coding 1502#4270.7$0.20$0.40262KOpen weights
23DeepSeek V4 Pro 0813DeepSeekArena coding 1503#3473.0$0.111$4.441MOpen weights
24Gemini 3.5 Flash LiteGoogleArena coding 1502#4171.1$0.30$2.501MNo linked weights
25Qwen3.7 PlusQwenArena coding 1500#4370.6$0.32$1.281MNo linked weights
26Qwen3.8 27BQwenArena coding 1498#6863.1$0.06$2.201MOpen weights
27GPT-5.4 MiniOpenAIArena coding 1495#4768.5$0.75$4.50400KNo linked weights
28InklingThinking MachinesArena coding 1493#6265.6$1.00$4.05524KOpen weights
29MiniMax M3MiniMaxArena coding 1492#6663.7$0.30$1.201MRestricted license
30MiMo-V2.5XiaomiArena coding 1491#6564.4$0.40$2.001MOpen weights
31Qwen3.5 397B A17BQwenArena coding 1490#5267.2$0.45$3.00262KOpen weights
32Grok 4.7xAIArena coding 1491 · Terminal-Bench 4.0 37.6%#5966.1$2.00$6.00500KNo linked weights
33Claude Haiku 5.5AnthropicArena coding 1487#9358.3$0.10$0.501MNo linked weights
34Inkling 2 SmallThinking MachinesArena coding 1486#9856.8$0.30$1.001MOpen weights
35Gemma 4 26B A4BGoogleArena coding 1483#6065.9$0.076$0.255262KOpen weights
36Kimi K2 ThinkingMoonshot AIArena coding 1486 · SWE-bench 63.4%#7162.3$0.60$2.50262KRestricted license
37Mistral Large 4Mistral AIArena coding 1482#9658.1$0.70$2.101MRestricted license
38Mistral Medium 3.5Mistral AIArena coding 1478#8759.4$1.50$7.50262KRestricted license
39DeepSeek V3.2DeepSeekArena coding 1476 · SWE-bench 70.0%#8958.8$0.26$0.38164KOpen weights
40Qwen3 235B A22B Instruct 2507QwenArena coding 1472#9158.5$0.14$0.80262KOpen weights
41Solar Pro 4UpstageArena coding 1461#14548.0$0.15$0.60524KNo linked weights
42R1 0528DeepSeekArena coding 1464#9558.1$0.571$2.29164KOpen weights
43Qwen3 VL 235B A22B InstructQwenArena coding 1463#10755.6$0.21$1.90262KOpen weights
44DeepSeek V3.1 TerminusDeepSeekArena coding 1459#9756.9$0.27$1.00164KOpen weights
45o3OpenAIArena coding 1460 · SWE-bench 58.4%#7561.0$2.00$8.00200KNo linked weights
46GPT-5.4 NanoOpenAIArena coding 1460#12651.6$0.20$1.25400KNo linked weights
47Qwen3.5-122B-A10BQwenArena coding 1458#9956.7$0.26$2.08262KOpen weights
48Qwen3 Coder 480B A35BQwenArena coding 1457 · SWE-bench 55.4%#14248.9$0.30$1.00262KOpen weights
49Qwen3 235B A22B Thinking 2507QwenArena coding 1442#11852.3$0.45$3.50128KOpen weights
50Qwen3 Next 80B A3B InstructQwenArena coding 1444#12451.7$0.10$1.10262KOpen weights
51Qwen3 235B A22BQwenArena coding 1444#11752.4——131KOpen weights
52Qwen3 30B A3B Instruct 2507QwenArena coding 1438#14747.7$0.10$0.30262KOpen weights
53GLM 4.5 AirZ.aiArena coding 1427#15445.1$0.13$0.85131KOpen weights
54GLM 4.6VZ.aiArena coding 1417#15146.1$0.30$0.90131KOpen weights
55Trinity Large ThinkingArcee AIArena coding 1413#15744.2——262KRestricted license
56Nemotron 3 SuperNVIDIAArena coding 1407#16341.7$0.075$0.40262KRestricted license
57Qwen3 32BQwenArena coding 1406#17539.5$0.08$0.28131KOpen weights
58Llama 3 3 Nemotron Super 49B v1 5NVIDIAArena coding 1404#17939.1———Restricted license
59granite 4.2 30bIBM GraniteArena coding 1398#19338.0——131KOpen weights
60gpt-oss-120bOpenAIArena coding 1389 · SWE-bench 26.0%#17439.8$0.15$0.60131KOpen weights
61Command ACohereArena coding 1388#16739.9$2.50$10.00256KGated access
62Llama 3 1 Nemotron Ultra 253B v1NVIDIAArena coding 1391#18039.1———Restricted license
63Qwen3 30B A3BQwenArena coding 1385#22234.6$0.12$0.5041KOpen weights
64QwQ 32BQwenArena coding 1382#19936.9———Open weights
65granite 4.2 3bIBM GraniteArena coding 1375#26826.5——131KOpen weights
66Llama 4 MaverickMetaArena coding 1372#22633.5$0.35$1.00524KGated access
67Granite 4.2 8BIBM GraniteArena coding 1368#26926.3$0.06$0.15131KOpen weights
68gpt-oss-20bOpenAIArena coding 1368#23832.2$0.018$0.09131KOpen weights
69Llama 4 ScoutMetaArena coding 1363#23931.8$0.10$0.301.3MGated access
70GPT-4o (2024-08-06)OpenAIArena coding 1360#19836.9$2.50$10.00128KNo linked weights
71Nemotron 3 Nano 30B A3BNVIDIAArena coding 1361#24531.2$0.06$0.24262KRestricted license
72Gemma 3 27BGoogleArena coding 1358#16042.7$0.08$0.16131KGated access
73Qwen2.5 72B InstructQwenArena coding 1355#25928.3$0.36$0.4033KRestricted license
74Llama 3.3 70B InstructMetaArena coding 1346#24431.4$0.71$0.71131KGated access
75Qwen2.5 Coder 32B InstructQwenArena coding 1342 · SWE-bench 9.0%#29023.3$0.66$1.0033KOpen weights
76Olmo 3.1 32B ThinkAllen InstituteArena coding 1339#27524.9———Open weights
77granite 4.0 h smallIBM GraniteArena coding 1330#27325.2———Open weights
78Gemma 3 12BGoogleArena coding 1316#20136.7$0.05$0.15131KGated access
79Llama 3 1 Nemotron 51B InstructNVIDIAArena coding 1312#27824.6———Restricted license
80gemma 3n E4B itGoogleArena coding 1307#24730.9——33KGated access
81Phi 4MicrosoftArena coding 1306#30819.2$0.07$0.1416KOpen weights
82Llama 3.1 Tulu 3 70BAllen InstituteArena coding 1308#28024.4———Restricted license
83Command R (08-2024)CohereArena coding 1281#31018.4$0.15$0.60128KGated access
84Gemma 3 4BGoogleArena coding 1273#26726.8$0.05$0.10131KGated access
85Command R+ (08-2024)CohereArena coding 1280#29422.9$2.50$10.00128KGated access
86Mixtral 8x22B InstructMistral AIArena coding 1278#31815.9$2.00$6.0066KOpen weights
87Qwen1.5 72B ChatQwenArena coding 1275#31716.1———Restricted license
88gemma 2 9b itGoogleArena coding 1270#30620.3———Gated access
89Qwen1.5 32B ChatQwenArena coding 1262#33313.0———Restricted license
90Llama 3.1 Tulu 3 8BAllen InstituteArena coding 1253#32014.9———Restricted license
91Llama 3.1 8B InstructMetaArena coding 1260#32513.8$0.02$0.04131KGated access
92Qwen1.5 14B ChatQwenArena coding 1240#35311.7———Restricted license
93Phi 3 medium 4k instructMicrosoftArena coding 1231#33912.2———Open weights
94gemma 1.1 7b itGoogleArena coding 1216#35410.5———Gated access
95Phi 3 small 8k instructMicrosoftArena coding 1205#3579.4———Open weights
96Qwen1.5 7B ChatQwenArena coding 1210#3607.3———Restricted license
97gemma 2 2b itGoogleArena coding 1191#35211.7———Gated access
98Mistral 7B Instruct v0.2Mistral AIArena coding 1186#3637.2———Open weights
99SOLAR 10.7B Instruct v1.0UpstageArena coding 1184#3597.6———Restricted license
100Llama 3.2 3B InstructMetaArena coding 1177#3588.5$0.051$0.335131KGated access

Best AI models for coding: common questions

Which model tops the best-for-coding list right now?

Gemini 4 Carbon (Google), as of 19 November 2026, with Arena coding 1567 on one public coding board and ModelCap Index position #1.

How is the best-for-coding list built?

It contains current, canonical language models with a measured public ModelCap Index position and admitted results on at least two public coding-family boards, ordered by the coding component of the ModelCap Index (ties broken by Index rank). A second tier, "Measured on one coding board", follows: Ranked models with a single admitted coding result, ordered by that board's normalised score. One board is a narrower claim than two, so these rows sit apart from the list above and never mix into it. Nothing is estimated for the list itself: the filter is a published fact and the order is a published figure, and every row links to the model page that shows its evidence.

How many models qualify for the best-for-coding list?

102 models qualify (0 in the primary tier, 102 measured on one board); the page shows the first 100. The population is the live ModelCap Index, so a model appears here the moment it holds a public position and meets the filter.

What is the cheapest model on the best-for-coding list?

Llama 3.1 8B Instruct lists the lowest blended API price at $0.025 per 1M tokens ($0.02 input / $0.04 output).

Which is the best open-weight model on this list?

MiMo-V2.6-Pro from Xiaomi is the highest-placed model with openly downloadable weights (#9 here, ModelCap Index #20).

How current is the best-for-coding list?

Positions, prices and context figures come from the same published snapshot dated 19 November 2026. Publication and page updates happen separately; use the snapshot date shown on the page.

Explore further