Token usage: output tokens per task
How many output tokens each configuration wrote per task of Artificial Analysis' Intelligence Index, reasoning included, averaged over the whole run. Fewer tokens for the same work means a cheaper, quicker model; it is not a measure of quality, so read it beside the index.
As of 9 Nov 2026, BenchLeader ranks GPT-6 Luna first on the Token usage: output tokens per task board at 3k, ahead of GPT-6.1 Terra (4k) and GPT-6.1 Sol (6k).
115 models · data as of 9 Nov 2026
Quality against what it costs to get there
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.
- 1GPT-6 Luna (low)3k
- 2GPT-6.1 Terra (low)4k
- 3GPT-6.1 Sol (low)6k
- 4GPT-6 Astra (low)9k
- 5Nemotron 3 Nano 4B12k
- 6Mistral Small 4 (thinking)14k
- 7Gemini 3.1 Flash Lite16k
- 8Claude Opus 5.5 (low)19k
- 9Claude Sonnet 5.5 (low)21k
- 10Claude Haiku 4.5 (thinking)25k
- 11Muse Glimmer (high)26k
- 12Claude Haiku 5.5 (low)28k
- 13North Mini Code30k
- 14Gemini 3.5 Flash Lite30k
- 15Beam 1.034k
| # | ||||||
|---|---|---|---|---|---|---|
| 1 | New Agnes 2.5 Pro Theta | 278k | 50.8 | 83% | 22 min | $0.575 |
| 2 |  Ling 3.1 Flashopen | 225k | 50.8 | 67% | – | $3.19 |
| 3 |  Qwen3.8 Maxmax | 171k | 58.9 | 70% | 60 min | $10.02 |
| 4 | NewNanbeige4.3-33Bopen | 158k | – | 64% | – | – |
| 5 | New Inkling 2 Small | 150k | 58.3 | 59% | 16 min | $1.18 |
| 6 |  MiniMax M3.1 Flashopen ↗ | 138k | 53.9 | 74% | 14 min | $0.491 |
| 7 |  GLM 5.3 Flashopen ↗ | 134k | 56.8 | 76% | 45 min | $0.583 |
| 8 |  MiMo-V2.6-Flashopen ↗ | 132k | 54.9 | 76% | 62 min | $0.129 |
| 9 | MiMo-V2.6-Proopen ↗ | 121k | 57.1 | 71% | 39 min | $0.257 |
| 10 |  Claude Opus 5max | 118k | 65.1 | 62% | 24 min | $11.82 |
| 11 |  Gemini 4 Argonhigh | 116k | 65.7 | 59% | 29 min | $9.95 |
| 12 |  Step 5 Previewopen | 113k | 54.6 | 70% | 14 min | $2.35 |
| 13 | MiniMax M3open ↗ | 91k | 56.3 | 50% | 12 min | $1.29 |
| 14 |  Muse Spark 1.3max | 91k | 61.7 | 60% | 12 min | $2.39 |
| 15 | Muse Wonder Fullmax | 88k | 62.2 | 62% | – | – |
| 16 | Gemini 4 Carbonlow | 86k | – | 38% | 22 min | $7.19 |
| 17 | Gemini 3.8 Flashmedium | 82k | 60.3 | 58% | 12 min | $6.26 |
| 18 |  Nemotron 3.5 Super VLopen ↗ | 81k | 45.8 | 69% | 5 min | $0.520 |
| 19 |  Hy4open ↗ | 78k | 55.3 | 70% | 14 min | $1.03 |
| 20 | NewMuse Emberxhigh | 72k | – | 61% | – | $0.116 |
| 21 |  Kimi K3.1 Proopen ↗ | 72k | 61.2 | 61% | 38 min | $3.68 |
| 22 |  DeepSeek V4 Prono reasoningopen ↗ | 72k | 42.3 | 0% | 17 min | $0.977 |
| 23 | Muse Wonderhigh | 68k | 56.3 | 57% | 9 min | $4.56 |
| 24 | Muse Spark Openxhighopen ↗ | 66k | 53.6 | 49% | 6 min | $2.02 |
| 25 |  GPT-5.6 Terramax | 66k | 60.5 | 62% | 10 min | $3.73 |
| 26 | DeepSeek V4.1 Flash 1017no reasoningopen ↗ | 65k | 45.0 | 0% | 8 min | $0.425 |
| 27 | Nemotron 3.5 Lightningopen ↗ | 65k | 41.5 | 47% | 4 min | $0.290 |
| 28 |  K2 Horizon 375B A23Bopen | 64k | 49.6 | 71% | 6 min | – |
| 29 | GPT-5.6 Lunamax | 59k | 56.2 | 64% | 9 min | $0.331 |
| 30 |  Grok 4.6xhigh | 58k | 57.6 | 55% | 13 min | $4.68 |
| 31 |  Mistral Large 4mediumopen ↗ | 56k | – | 63% | 9 min | $1.41 |
| 32 |  K Exaone 2.0 0803open | 55k | 42.4 | 41% | – | – |
| 33 | GPT-6 Solmax | 54k | 62.9 | 66% | 7 min | $2.47 |
| 34 | GLM 5.3lowopen ↗ | 52k | 47.4 | 69% | 16 min | $1.57 |
| 35 |  Granite 4.2 8Bopen ↗ | 52k | 41.1 | 72% | 9 min | $0.040 |
| 36 | Nemotron 3 Ultra 550B A55Bthinkingopen ↗ | 51k | 49.2 | 40% | 5 min | $2.23 |
| 37 | Inklingxhighopen ↗ | 51k | 50.1 | 66% | 8 min | $1.35 |
| 38 | Qwen3.8 27Bno reasoningopen ↗ | 45k | 45.6 | 0% | 19 min | $3.77 |
| 39 | Grok 4.7low | 43k | 55.5 | 51% | 11 min | $2.92 |
| 40 | Kimi K3lowopen ↗ | 40k | 53.7 | 68% | 18 min | $2.37 |
| 41 | Mistral Medium 3.5open | 40k | 46.6 | 45% | 5 min | $1.04 |
| 42 | Claude Fable 5.1low | 38k | 58.7 | 41% | 7 min | $4.47 |
| 43 | Inkling Smallopen ↗ | 37k | 52.1 | 68% | 8 min | $0.203 |
| 44 | NewClaude Fable 5.2low | 36k | 65.3 | 46% | – | $3.10 |
| 45 |  Command A+open | 36k | 45.9 | 57% | 3 min | – |
| 46 | Granite 4.2 3Bopen ↗ | 34k | 39.4 | 56% | – | – |
| 47 |  Beam 1.0open | 34k | 45.6 | 64% | 2 min | $0.284 |
| 48 | Gemini 3.5 Flash Lite | 30k | 51.5 | 55% | 102 s | $0.693 |
| 49 | North Mini Codeopen | 30k | 40.2 | 57% | 3 min | – |
| 50 | Claude Haiku 5.5low | 28k | 47.8 | 54% | 2 min | $0.044 |
| 51 | Muse Glimmerhighopen ↗ | 26k | 46.6 | 55% | 5 min | $0.170 |
| 52 | Claude Haiku 4.5thinking | 25k | 45.3 | 17% | 4 min | $0.532 |
| 53 | Claude Sonnet 5.5low | 21k | 50.8 | 41% | 3 min | $0.511 |
| 54 | Claude Opus 5.5low | 19k | 59.0 | 36% | 3 min | $0.976 |
| 55 | Gemini 3.1 Flash Lite | 16k | 52.4 | 68% | 50 s | $0.194 |
| 56 | Mistral Small 4thinkingopen | 14k | 40.5 | 29% | 99 s | $0.040 |
| 57 | Nemotron 3 Nano 4Bopen | 12k | 37.8 | 81% | – | – |
| 58 | GPT-6 Astralow | 9k | 62.2 | 36% | 4 min | $1.42 |
| 59 | GPT-6.1 Sollow | 6k | 60.2 | 27% | 105 s | $0.161 |
| 60 | GPT-6.1 Terralow | 4k | 56.7 | 21% | 64 s | $0.071 |
Cite as: BenchLeader, “Token usage: output tokens per task”, data as of 9 Nov 2026, https://www.benchleader.com/leaderboards/tokens-per-task.