Token efficiency: quality per dollar of real work
BenchLeader Index divided by what it cost Artificial Analysis to run its whole Intelligence Index on that configuration. Unlike price per token this counts the tokens a model actually burns, so a model that thinks at length scores lower than its sticker price suggests.
As of 21 Sept 2026, BenchLeader ranks Granite 4.2 3B first on the Token efficiency: quality per dollar of real work board at 7625, ahead of GPT-5.6 Luna (5310) and GPT-5.6 Luna (4549).
128 models · data as of 21 Sept 2026
Quality against what it costs to get there
OpenAI31
Anthropic18
Alibaba15
Google12
Mistral AI8
SpaceXAI7
DeepSeek6
Meta5
Zhipu AI5
Moonshot AI4
NVIDIA4
IBM2
MiniMax2
Xiaomi2
Ant Group1
Arcee AI1- Celeris1
Meituan1
The dashed staircase is the efficiency frontier: at each price, the best index anyone reaches. A model above and left of the rest gives you more quality per dollar of real work. Size bands and run costs come from Artificial Analysis, which is the only source that publishes what a fixed batch of work actually cost.
- 1Granite 4.2 3B7625
- 2GPT-5.6 Luna (low)5310
- 3gpt-oss-20b (high)3945
- 4Nemotron 3 Nano 30B A3B (thinking)2812
- 5DeepSeek V32268
- 6Granite 4.2 8B1983
- 7Gemini 3.1 Flash Lite1383
- 8Mistral Small 3.11218
- 9MiMo-V2.5-Pro1093
- 10GPT-5 mini (high)1032
- 11Mistral Small 4 (thinking)1027
- 12Muse Glimmer (high)958
- 13Ministral 3 3B887
- 14LongCat 2.0853
- 15Celeris-1835
| # | |||||
|---|---|---|---|---|---|
| 1 | Granite 4.2 3BIBMopen ↗ | 7625 | 45.6 | $0.0060 | – |
| 2 | GPT-5.6 LunalowOpenAI | 5310 | 52.2 | $0.0098 | – |
| 3 | gpt-oss-20bhighOpenAIopen ↗ | 3945 | 44.7 | $0.011 | – |
| 4 | Nemotron 3 Nano 30B A3BthinkingNVIDIAopen ↗ | 2812 | 47.1 | $0.017 | – |
| 5 | DeepSeek V3DeepSeekopen ↗ | 2268 | 44.8 | $0.020 | – |
| 6 | Granite 4.2 8BIBMopen ↗ | 1983 | 46.8 | $0.024 | – |
| 7 | Gemini 3.1 Flash LiteGoogle | 1383 | 54.5 | $0.039 | – |
| 8 | Mistral Small 3.1Mistral AIopen | 1218 | 39.9 | $0.033 | – |
| 9 | MiMo-V2.5-ProXiaomiopen ↗ | 1093 | 59.1 | $0.054 | – |
| 10 | GPT-5 minihighOpenAI | 1032 | 55.2 | $0.053 | – |
| 11 | Mistral Small 4thinkingMistral AIopen | 1027 | 46.3 | $0.045 | – |
| 12 | Muse GlimmerhighMetaopen ↗ | 958 | 52.8 | $0.055 | – |
| 13 | Ministral 3 3BMistral AIopen | 887 | 37.5 | $0.042 | – |
| 14 | LongCat 2.0Meituanopen ↗ | 853 | 50.3 | $0.059 | – |
| 15 | Celeris-1Celeris | 835 | 41.9 | $0.050 | – |
| 16 | Qwen3 235B A22B 2507thinkingAlibabaopen ↗ | 634 | 51.1 | $0.081 | – |
| 17 | Ministral 3 8BMistral AIopen | 594 | 38.9 | $0.066 | – |
| 18 | Qwen3 30B A3B 2507thinkingAlibabaopen ↗ | 576 | 44.1 | $0.077 | – |
| 19 |  Solar Pro 3Upstage | 553 | 43.9 | $0.079 | – |
| 20 | MiniMax M2.7MiniMaxopen ↗ | 551 | 56.0 | $0.102 | – |
| 21 | Nemotron 3.5 LightningNVIDIAopen ↗ | 518 | 48.2 | $0.093 | – |
| 22 | gpt-oss-120bhighOpenAIopen ↗ | 462 | 49.6 | $0.107 | – |
| 23 | Gemini 3.5 Flash LiteGoogle | 449 | 55.4 | $0.124 | – |
| 24 | Mistral Large 3Mistral AIopen | 441 | 46.0 | $0.104 | – |
| 25 | GPT-5.6 TerralowOpenAI | 400 | 57.8 | $0.144 | – |
| 26 | Grok 4.3highSpaceXAI | 352 | 58.1 | $0.165 | – |
| 27 | GPT-5.4 NanoxhighOpenAI | 298 | 54.5 | $0.183 | – |
| 28 | Ministral 3 14BMistral AIopen | 287 | 40.0 | $0.139 | – |
| 29 | DeepSeek V4 FlashmaxDeepSeekopen ↗ | 282 | 62.0 | $0.220 | – |
| 30 | GLM 5.3 FlashZhipu AIopen ↗ | 252 | 63.9 | $0.253 | – |
| 31 | GPT-5.6 SollowOpenAI | 244 | 63.6 | $0.261 | – |
| 32 | Gemini 2.5 ProGoogle | 235 | 54.4 | $0.232 | – |
| 33 | DeepSeek V4.1 FlashmaxDeepSeekopen ↗ | 233 | 61.8 | $0.265 | – |
| 34 | Claude Haiku 4.5thinkingAnthropic | 228 | 47.4 | $0.208 | – |
| 35 | Deepseek v4 Flash VisionmaxDeepSeek | 190 | 59.7 | $0.314 | – |
| 36 | DeepSeek R1DeepSeekopen ↗ | 186 | 48.7 | $0.262 | – |
| 37 | Qwen3.7 PlusAlibaba | 184 | 59.8 | $0.325 | – |
| 38 | Ring-2.6-1TAnt Groupopen | 174 | 50.0 | $0.287 | – |
| 39 | Qwen3.8-Flash-NextAlibabaopen ↗ | 165 | 61.5 | $0.372 | – |
| 40 | Qwen3.5-122B-A10BthinkingAlibabaopen ↗ | 164 | 53.4 | $0.325 | – |
| 41 | GPT-5.4 MinixhighOpenAI | 136 | 55.7 | $0.410 | – |
| 42 | Trinity LargethinkingArcee AIopen | 135 | 47.6 | $0.353 | – |
| 43 | Grok 4.6lowSpaceXAI | 128 | 60.9 | $0.475 | – |
| 44 | Qwen3.5 397B A17BthinkingAlibabaopen ↗ | 118 | 56.0 | $0.475 | – |
| 45 | Mistral Medium 3.5Mistral AIopen | 117 | 51.2 | $0.437 | – |
| 46 | Mistral Medium 3Mistral AI | 116 | 45.5 | $0.392 | – |
| 47 | MiniMax M3MiniMaxopen ↗ | 113 | 57.4 | $0.508 | – |
| 48 | Qwen3.6 35B A3BthinkingAlibabaopen ↗ | 112 | 53.0 | $0.475 | – |
| 49 | Kimi K2.7 CodeMoonshot AIopen ↗ | 103 | 55.8 | $0.541 | – |
| 50 | Nemotron 3 Ultra 550B A55BthinkingNVIDIAopen ↗ | 99.3 | 57.5 | $0.579 | – |
| 51 | Claude Sonnet 4.5thinkingAnthropic | 95.8 | 53.3 | $0.557 | – |
| 52 | DeepSeek V4 PromaxDeepSeekopen ↗ | 95.1 | 64.1 | $0.674 | – |
| 53 | Gemini 3.1 ProGoogle | 94.6 | 63.8 | $0.675 | – |
| 54 | New Step 5 PreviewStepFun | 90.7 | 64.9 | $0.716 | – |
| 55 | Qwen3.6 27BthinkingAlibabaopen ↗ | 88.2 | 54.8 | $0.621 | – |
| 56 | GPT-5.5 InstantOpenAI | 85.3 | 59.0 | $0.691 | – |
| 57 | GPT-6 AstralowOpenAI | 83.4 | 68.2 | $0.818 | – |
| 58 | Qwen3 Coder NextAlibabaopen ↗ | 77.0 | 42.5 | $0.552 | – |
| 59 | Kimi K2.6Moonshot AIopen ↗ | 75.5 | 60.4 | $0.800 | – |
| 60 | Qwen3.8 27BxhighAlibabaopen ↗ | 72.1 | 59.1 | $0.820 | – |
Cite as: BenchLeader, “Token efficiency: quality per dollar of real work”, https://www.benchleader.com/leaderboards/efficiency, data as of 21 Sept 2026.