BenchLeader
Menu
Best LLMsChoose a modelCoding plansLeaderboardsModelsBenchmarksProvidersReleases12

Token efficiency: quality per dollar of real work

BenchLeader Index divided by what an average task of Artificial Analysis' Intelligence Index cost 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 9 Nov 2026, BenchLeader ranks GPT-6 Luna first on the Token efficiency: quality per dollar of real work board at 8945, ahead of Claude Haiku 5.5 (1092) and Granite 4.2 8B (1016).

108 models · data as of 9 Nov 2026

Quality against what it costs to get there

Model size
40506070$0.0050$0.010$0.020$0.050$0.100$0.200$0.500$1.00$2.00$5.00$10.00$20.00What an average Intelligence Index task cost ($, log scale)BenchLeader Index v1.4Claude Fable 5.2Claude Opus 5.5Claude Fable 5.2NEW Claude Fable 5.2Claude Fable 5.2Gemini 4 CarbonGemini 4 CarbonGPT-6 AstraGPT-6.1 SolGPT-6.1 SolClaude Fable 5.1GPT-6 AstraGPT-6 AstraClaude Fable 5.2GPT-6 AstraGemini 4 ArgonClaude Opus 5.5Claude Opus 5.5Claude Fable 5.2Claude Opus 5Claude Opus 5Claude Sonnet 5.5GPT-6.1 TerraGPT-6.1 TerraClaude Fable 5.1GPT-6 AstraGemini 4 CarbonGPT-6.1 SolGPT-6.1 SolClaude Fable 5.1Claude Fable 5.1GPT-6 SolGPT-6.1 SolClaude Sonnet 5.5GPT-6 AstraMuse WonderGemini 3.8 FlashGemini 3.8 FlashMuse Spark 1.3Claude Fable 5.1Claude Opus 5.5Kimi K3.1 ProGPT-5.6 TerraGPT-5.6 TerraGemini 3.8 FlashGPT-6.1 SolGPT-6.1 TerraGPT-6.1 TerraClaude Opus 5.5GLM 5.3Qwen3.8 MaxQwen3.8 MaxDeepSeek V4.1 Flash 1017Claude Fable 5.1GPT-6.1 TerraInkling 2 SmallGrok 4.7Claude Sonnet 5.5Grok 4.6Grok 4.6MiMo-V2.6-ProGrok 4.7Grok 4.7DeepSeek V4 ProGLM 5.3 FlashGLM 5.3 FlashClaude Haiku 5.5Claude Haiku 5.5GPT-6.1 TerraGPT-6 LunaMiniMax M3Muse WonderMuse WonderGPT-5.6 LunaClaude Haiku 5.5Grok 4.7Hy4Hy4MiMo-V2.6-FlashMiMo-V2.6-FlashStep 5 PreviewGLM 5.3 FlashMiniMax M3.1 FlashMiniMax M3.1 FlashKimi K3Kimi K3Muse Spark OpenMuse Spark OpenGemini 3.1 Flash LiteQwen3.8 27BQwen3.8 27BMistral Large 4Mistral Large 4Inkling SmallGPT-6 LunaGPT-6.1 TerraGemini 3.5 Flash LiteGPT-6 LunaDeepSeek V4.1 Flash 1017Claude Sonnet 5.5Claude Sonnet 5.5Claude Sonnet 5.5Agnes 2.5 Pro ThetaLing 3.1 FlashLing 3.1 FlashQwen3.8 27BClaude Haiku 5.5Claude Haiku 5.5InklingInklingGPT-6 LunaGPT-6 LunaMistral Large 4Nemotron 3 Ultra 550B A55BNemotron 3 Ultra 550B A55BQwen3.8 27BGemini 3.1 Flash LiteGemini 3.5 Flash LiteClaude Haiku 5.5GLM 5.3GLM 5.3Inkling SmallInkling SmallMistral Medium 3.5Mistral Medium 3.5Muse GlimmerMuse GlimmerGPT-6 LunaNemotron 3.5 Super VLBeam 1.0Beam 1.0Qwen3.8 27BClaude Haiku 4.5DeepSeek V4.1 Flash 1017DeepSeek V4 ProDeepSeek V4 ProMistral Medium 3.5Nemotron 3.5 LightningNemotron 3.5 LightningGranite 4.2 8BGranite 4.2 8BMistral Small 4Mistral Small 4GPT-6 Luna
Purple: the frontier, the models no other model shown beats on both axes. Dashed ring: new in the last seven days.

The purple line 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, task costs and token counts come from Artificial Analysis, which is the only source that publishes what a fixed batch of work actually cost and how much a model wrote for it.

Top 15 by index per $
  1. 1GPT-6 Luna (low)8945
  2. 2Claude Haiku 5.5 (low)1092
  3. 3Granite 4.2 8B1016
  4. 4Mistral Small 4 (thinking)1015
  5. 5GPT-6.1 Terra (low)795
  6. 6MiMo-V2.6-Flash427
  7. 7GPT-6.1 Sol (low)373
  8. 8Muse Glimmer (high)275
  9. 9Gemini 3.1 Flash Lite270
  10. 10Inkling Small256
  11. 11MiMo-V2.6-Pro222
  12. 12GPT-5.6 Luna (max)170
  13. 13Beam 1.0160
  14. 14Nemotron 3.5 Lightning143
  15. 15DeepSeek V4.1 Flash 1017 (high)112
52 of 52
#
1GPT-6 Lunalow894545.8$0.0051–
2Claude Haiku 5.5low109247.8$0.044–
3Granite 4.2 8B101641.1$0.040Small
4Mistral Small 4thinking101540.5$0.040Medium
5GPT-6.1 Terralow79556.7$0.071–
6MiMo-V2.6-Flash42754.9$0.129Large
7GPT-6.1 Sollow37360.2$0.161–
8Muse Glimmerhigh27546.6$0.170Small
9Gemini 3.1 Flash Lite27052.4$0.194–
10Inkling Small25652.1$0.203Large
11MiMo-V2.6-Pro22257.1$0.257Large
12GPT-5.6 Lunamax17056.2$0.331–
13Beam 1.016045.6$0.284Large
14Nemotron 3.5 Lightning14341.5$0.290Small
15DeepSeek V4.1 Flash 1017high11251.2$0.456Large
16MiniMax M3.1 Flash11053.9$0.491Large
17Claude Sonnet 5.5low99.350.8$0.511–
18GLM 5.3 Flash97.456.8$0.583Large
19NewAgnes 2.5 Pro Theta88.450.8$0.575–
20Nemotron 3.5 Super VL88.145.8$0.520Medium
21Claude Haiku 4.5thinking85.245.3$0.532–
22Gemini 3.5 Flash Lite74.351.5$0.693–
23Claude Opus 5.5low60.559.0$0.976–
24Hy453.655.3$1.03–
25NewInkling 2 Small49.558.3$1.18Large
26Mistral Medium 3.544.846.6$1.04Medium
27MiniMax M343.856.3$1.29Large
28GPT-6 Astralow43.762.2$1.42–
29DeepSeek V4 Prono reasoning43.342.3$0.977Large
30Inklingxhigh37.150.1$1.35Large
31GLM 5.3low30.147.4$1.57Large
32Muse Spark Openxhigh26.653.6$2.02Large
33Muse Spark 1.3max25.961.7$2.39–
34Qwen3.8 27Blow25.749.0$1.91Small
35GPT-6 Solmax25.562.9$2.47–
36Mistral Large 4high23.949.3$2.06Large
37Step 5 Preview23.354.6$2.35Large
38Kimi K3low22.753.7$2.37Large
39Nemotron 3 Ultra 550B A55Bthinking22.149.2$2.23Large
40NewClaude Fable 5.2low21.165.3$3.10–
41Grok 4.7low19.055.5$2.92–
42Kimi K3.1 Pro16.661.2$3.68Large
43GPT-5.6 Terramax16.260.5$3.73–
44Ling 3.1 Flash15.950.8$3.19Large
45Claude Fable 5.1low13.158.7$4.47–
46Muse Wonderhigh12.356.3$4.56–
47Grok 4.6xhigh12.357.6$4.68–
48Gemini 4 Carbonmedium9.764.1$6.59–
49Gemini 3.8 Flashmedium9.660.3$6.26–
50Gemini 4 Argonhigh6.665.7$9.95–
51Qwen3.8 Maxmax5.958.9$10.02–
52Claude Opus 5max5.565.1$11.82–

Cite as: BenchLeader, “Token efficiency: quality per dollar of real work”, data as of 9 Nov 2026, https://www.benchleader.com/leaderboards/efficiency.