Models / API speed
LLM API speed
How long each model takes on a real task, how fast it writes, and how much it writes to get there. Seconds per task are measured on our own benchmarks; tokens per second are from Artificial Analysis. The pricing page does the same for cost.
- Benchmarks timed
- 4
- seconds per task, from our runs
- Settings with tokens/s
- 137
- model and reasoning setting
- Speeds measured
- 9 Nov 2026
Seconds per task
Wall-clock time from sending the request to a usable answer, measured while running our own benchmarks, with the score the same run earned. It's what a model's speed means in practice: how long it thinks, how much it writes and how fast it writes, together. Each model is shown at its quickest setting. Times were measured through OpenRouter on the run date, so they also depend on the host and its load that day.
CatalogBench: turning a sparse product feed into a listing (36 models)
CatalogBench: score against time per listing
Further up scored better; further left finished sooner. The line joins the models no other model beats on both.
Best score for the waitOther models (hover for names)Best score at each wait
| # | Model and setting | Seconds per listing | Output tokens per listing | Reliably publish-ready |
|---|---|---|---|---|
| 1 | Gemini 3.5 Flash-Lite | 3.3 s | not yet | 17.9% |
| 2 | Mistral Medium 3.5 | 4.9 s | not yet | 17.9% |
| 3 | GPT-6.2 Sol | 6.5 s | 442 | 75.0% |
| 4 | DeepSeek-V4.1-Flash-1017 | 6.7 s | 1,226 | 21.4% |
| 5 | Gemini 4 Argon | 7.3 s | not yet | 60.7% |
| 6 | Claude Haiku 4.5 | 8.1 s | not yet | 1.8% |
| 7 | Muse Ember | 8.1 s | 3,816 | 53.6% |
| 8 | GPT-6 Luna | 8.2 s | not yet | 55.4% |
| 9 | Mistral Large 4 | 8.2 s | not yet | 7.1% |
| 10 | GPT-6.2 Astra | 9.7 s | 426 | 73.2% |
| 11 | Claude Sonnet 5.5 | 9.9 s | not yet | 32.1% |
| 12 | GPT-6.1 Terra | 10 s | 457 | 64.3% |
| 13 | GPT-6.1 Sol | 10 s | not yet | 75.0% |
| 14 | Muse Wonder | 10 s | not yet | 66.1% |
| 15 | GPT-6 Sol | 11 s | not yet | 66.1% |
| 16 | Claude Haiku 5.5 | 11 s | not yet | 26.8% |
| 17 | GPT-6 Astra | 12 s | not yet | 75.0% |
| 18 | Claude Opus 5.5 | 17 s | not yet | 55.4% |
| 19 | Claude Fable 5.2 | 18 s | 1,442 | 71.4% |
| 20 | DeepSeek-V4.1-Flash | 18 s | not yet | 41.1% |
| 21 | Gemini 3.8 Flash | 19 s | not yet | 35.7% |
| 22 | Gemini 4 Hydrogen | 20 s | 2,416 | 62.5% |
| 23 | Qwen4-Max | 21 s | 1,227 | 39.3% |
| 24 | Claude Fable 5.1 | 22 s | not yet | 26.8% |
| 25 | Gemini 3.1 Pro Preview | 22 s | not yet | 44.6% |
| 26 | Particle-1 | 24 s | 2,180 | 0.0% |
| 27 | MiniMax M3.1 Flash | 25 s | 3,014 | 0.0% |
| 28 | GLM-5.3 FlashX Legal | 29 s | 3,651 | 0.0% |
| 29 | Gemini 4 Neon | 36 s | 3,939 | 78.6% |
| 30 | Grok 4.7 | 42 s | not yet | 57.1% |
| 31 | Gemini 4 Carbon | 45 s | 1,600 | 53.6% |
| 32 | Muse Spark 1.3 | 48 s | not yet | 48.2% |
| 33 | GLM-5V-Turbo | 62 s | not yet | 1.8% |
| 34 | Kimi K3 | 67 s | not yet | 41.1% |
| 35 | Qwen3.8-Max (0902) | 78 s | not yet | 28.6% |
| 36 | Kimi K3.1 Pro | 114 s | 3,996 | – |
DeckBench: turning an analysis into a deck (36 models)
DeckBench: score against time per deck
Further up scored better; further left finished sooner. The line joins the models no other model beats on both.
Best score for the waitOther models (hover for names)Best score at each wait
| # | Model and setting | Seconds per deck | Output tokens per deck | Deck rating |
|---|---|---|---|---|
| 1 | Gemini 3.5 Flash-Lite | 15 s | not yet | 605 |
| 2 | Mistral Medium 3.5 | 25 s | not yet | 585 |
| 3 | Muse Wonder | 29 s | not yet | 1,204 |
| 4 | Muse Ember | 30 s | 22,094 | 1,143 |
| 5 | GPT-6.2 Astra | 44 s | 4,453 | 1,188 |
| 6 | Claude Haiku 4.5 | 52 s | not yet | 646 |
| 7 | Claude Haiku 5.5 | 62 s | not yet | 1,120 |
| 8 | Gemini 4 Argon | 68 s | not yet | 951 |
| 9 | DeepSeek-V4.1-Flash-1017 | 73 s | 21,019 | 1,046 |
| 10 | Gemini 3.1 Pro Preview | 78 s | not yet | 716 |
| 11 | GPT-6.2 Sol | 82 s | 7,306 | 1,370 |
| 12 | Mercury 3 | 83 s | 32,448 | 726 |
| 13 | Claude Sonnet 5.5 | 83 s | not yet | 1,334 |
| 14 | Muse Spark 1.3 | 89 s | not yet | 807 |
| 15 | Gemini 3.8 Flash | 93 s | not yet | 771 |
| 16 | Claude Opus 5.5 | 98 s | not yet | 1,323 |
| 17 | GPT-6 Luna | 101 s | not yet | 1,252 |
| 18 | Mistral Large 4 | 114 s | not yet | 595 |
| 19 | GPT-6.1 Terra | 2.2 min | 12,086 | 1,280 |
| 20 | Particle-1 | 2.6 min | 16,952 | 1,086 |
| 21 | Claude Fable 5.1 | 2.6 min | not yet | 1,256 |
| 22 | GPT-6 Sol | 2.8 min | not yet | 1,438 |
| 23 | GLM-5.3 FlashX Legal | 2.9 min | 27,796 | 809 |
| 24 | GPT-6.1 Sol | 3.0 min | not yet | 1,375 |
| 25 | Gemini 4 Hydrogen | 3.4 min | 31,368 | 1,079 |
| 26 | MiniMax M3.1 Flash | 3.6 min | 33,765 | 1,266 |
| 27 | Gemini 4 Carbon | 3.8 min | 15,626 | 992 |
| 28 | Claude Fable 5.2 | 4.8 min | 33,285 | 1,403 |
| 29 | GPT-6 Astra | 5.2 min | not yet | 1,551 |
| 30 | DeepSeek-V4-Pro (0423) | 5.5 min | not yet | 530 |
| 31 | Grok 4.7 | 6.2 min | not yet | 1,242 |
| 32 | Gemini 4 Neon | 7.1 min | 48,462 | 1,013 |
| 33 | Kimi K3 | 7.3 min | not yet | 1,194 |
| 34 | Qwen4-Max | 8.1 min | 36,572 | 1,563 |
| 35 | Qwen3.8-Max (0902) | 9.1 min | not yet | 1,025 |
| 36 | Kimi K3.1 Pro | 21 min | 51,739 | 1,223 |
SurveyBench: planning a user survey and analysing its responses (27 models)
SurveyBench: score against time per survey (both steps)
Further up scored better; further left finished sooner. The line joins the models no other model beats on both.
Best score for the waitOther models (hover for names)Best score at each wait
| # | Model and setting | Seconds per survey (both steps) | Output tokens per survey (both steps) | Analysis rating |
|---|---|---|---|---|
| 1 | Gemini 3.5 Flash-Lite | 16 s | not yet | 637 |
| 2 | Mistral Medium 3.5 | 48 s | not yet | 578 |
| 3 | Muse Ember | 73 s | 40,298 | 1,344 |
| 4 | Gemini 4 Argon | 74 s | not yet | 928 |
| 5 | Muse Wonder | 83 s | not yet | 1,439 |
| 6 | GPT-6.2 Astra | 92 s | 5,888 | 1,476 |
| 7 | GPT-6.1 Terra | 92 s | 8,112 | 1,217 |
| 8 | Claude Haiku 4.5 | 93 s | not yet | 560 |
| 9 | GPT-6.2 Sol | 104 s | 9,187 | 1,362 |
| 10 | Gemini 3.8 Flash | 2.2 min | 21,981 | 845 |
| 11 | DeepSeek-V4.1-Flash-1017 | 2.2 min | 30,893 | 1,119 |
| 12 | GPT-6 Luna | 2.3 min | 14,333 | 1,072 |
| 13 | Claude Haiku 5.5 | 2.5 min | not yet | 1,233 |
| 14 | Gemini 3.1 Pro Preview | 2.8 min | not yet | 744 |
| 15 | Claude Opus 5.5 | 3.0 min | not yet | 1,417 |
| 16 | Mercury 3 | 3.2 min | 57,343 | 735 |
| 17 | GPT-6.1 Sol | 3.7 min | not yet | 1,429 |
| 18 | Particle-1 | 4.2 min | 26,601 | 1,133 |
| 19 | Gemini 4 Hydrogen | 4.4 min | 41,542 | 1,096 |
| 20 | Claude Fable 5.2 | 5.6 min | 33,902 | 1,670 |
| 21 | MiniMax M3.1 Flash | 5.9 min | 63,237 | 428 |
| 22 | Gemini 4 Carbon | 6.5 min | 27,760 | 1,054 |
| 23 | GLM-5.3 FlashX Legal | 7.2 min | 51,553 | 937 |
| 24 | Gemini 4 Neon | 7.3 min | 50,889 | 629 |
| 25 | Mistral Large 4 | 7.3 min | not yet | 770 |
| 26 | Qwen4-Max | 12 min | 39,999 | 1,381 |
| 27 | Kimi K3.1 Pro | 33 min | 70,935 | 605 |
BulletBench: choosing a chess move on the Lightning 10+1 clock (45 models)
BulletBench: score against time per move (median)
Further up scored better; further left finished sooner. The line joins the models no other model beats on both.
Best score for the waitOther models (hover for names)Best score at each wait
Tokens per task. Some models solve a task in far fewer tokens than others, which can matter more than how fast they write. We record output tokens per task, thinking included, for every run from 14 October 2026; a benchmark shows them once its models have been run again. Earlier runs recorded only time and cost.
Tokens per second
How fast each model writes once it has started: the median output tokens a second on its usual API, measured by Artificial Analysis. A reasoning model thinks before it answers, and that wait isn't in tokens per second, so "first token" shows it: the median time before the answer starts. Each row is one model at one reasoning setting. Models with a shutdown date announced are left out.
No model matches . Try part of a name or a developer, such as "gemini", or .
| Model and setting | Tokens/s | First token | Intelligence |
|---|---|---|---|
| Nemotron 3.5 Super VLNVIDIA · open weights | 411 | 5.4 s | 26.5 |
| Gemini 3.5 Flash-LiteGoogle | 398 | 4.7 s | 22.2 |
| DeepSeek-V4.1-Flash (no reasoning)DeepSeek · open weights | 369 | 0.6 s | 24.7 |
| Trinity Large ThinkingArcee AI | 362 | 6.6 s | 10.8 |
| Beam 1.0 501B A23BReflection · open weights | 323 | 6.7 s | 30.3 |
| gpt-oss-20b (low reasoning)OpenAI · open weights | 277 | 7.6 s | 10.0 |
| gpt-oss-20b (high reasoning)OpenAI · open weights | 263 | 8.0 s | 9.0 |
| Nova Micro 1.0Amazon | 260 | 0.5 s | 5.9 |
| gpt-oss-120b (high reasoning)OpenAI · open weights | 259 | 8.2 s | 11.6 |
| gpt-oss-120b (low reasoning)OpenAI · open weights | 258 | 8.3 s | 10.2 |
| Muse Spark 1.3 (xhigh reasoning)Meta | 254 | 30 s | 45.1 |
| Claude Haiku 5.5 (max reasoning)Anthropic | 240 | 4.7 min | 43.4 |
| Nemotron 3 Nano 30B A3B (thinking reasoning)NVIDIA · open weights | 238 | 8.9 s | 8.9 |
| Nova 2 Lite (no reasoning)Amazon | 233 | 0.7 s | 8.7 |
| Inkling (xhigh reasoning)Thinking Machines · open weights | 232 | 11 s | 25.0 |
| Nemotron 3 Nano 30B A3B (no reasoning)NVIDIA · open weights | 232 | 0.4 s | 6.8 |
| DeepSeek-V4.1-Flash (max reasoning)DeepSeek · open weights | 231 | 9.4 s | 39.5 |
| Qwen3.5-Omni-FlashAlibaba | 231 | 0.9 s | 12.5 |
| Nova 2 Lite (high reasoning)Amazon | 220 | 24 s | 13.4 |
| DeepSeek-V4.1-Flash-1017 (max reasoning)DeepSeek · open weights | 212 | 10 s | 45.1 |
| Ministral 3 3BMistral AI · open weights | 209 | 0.4 s | 4.8 |
| Nova 2 Lite (medium reasoning)Amazon | 200 | 27 s | 12.5 |
| Claude Haiku 5.5 (xhigh reasoning)Anthropic | 192 | 46 s | 41.2 |
| DeepSeek-V4-Flash-Vision-Exp (max reasoning)DeepSeek · open weights | 191 | 12 s | 34.8 |
| Inkling SmallThinking Machines · open weights | 181 | 13 s | 25.7 |
| Nova 2 Lite (low reasoning)Amazon | 177 | 16 s | 11.8 |
| Mistral Small 4Mistral AI · open weights | 176 | 12 s | 11.3 |
| Claude Haiku 5.5 (high reasoning)Anthropic | 169 | 16 s | 37.8 |
| Mistral Small 4 (no reasoning)Mistral AI · open weights | 166 | 0.6 s | 9.0 |
| Qwen3-Next-80B-A3B-ThinkingAlibaba | 163 | 13 s | 11.2 |
| MiniMax M3.1 FlashMiniMax · open weights | 159 | 13 s | 42.4 |
| Mistral Medium 3.5Mistral AI | 157 | 13 s | 14.2 |
| Claude Haiku 5.5 (low reasoning)Anthropic | 157 | 6.4 s | 29.5 |
| Claude Haiku 5.5 (medium reasoning)Anthropic | 155 | 8.9 s | 34.5 |
| o3OpenAI | 155 | 4.6 s | 20.2 |
| Muse Wonder (max reasoning)Meta | 153 | 48 s | 52.4 |
| GPT-5.5 Instant (2026-06-26)OpenAI | 152 | 14 s | 26.0 |
| Gemma 4 12BGoogle | 148 | 15 s | 14.2 |
| MiniMax-M3MiniMax · open weights | 146 | 15 s | 29.2 |
| Gemma 4 12B (no reasoning)Google | 145 | 1.1 s | 9.4 |
| Qwen3.5-122B-A10B (no reasoning)Alibaba · open weights | 144 | 1.0 s | 17.7 |
| Muse Glimmer (high reasoning)Meta | 142 | 14 s | 17.5 |
| Muse Spark Open (xhigh reasoning)Meta · open weights | 142 | 15 s | 38.9 |
| Gemini 3.8 Flash (high reasoning)Google | 141 | 16 s | 40.9 |
| GPT-5.6 Terra (max reasoning)OpenAI | 140 | 3.3 min | 42.1 |
| Claude Sonnet 5.5 (max reasoning)Anthropic | 137 | 5.4 min | 56.0 |
| Qwen3.6-35B-A3B (no reasoning)Alibaba · open weights | 135 | 1.1 s | 15.2 |
| Qwen3.6-35B-A3BAlibaba · open weights | 134 | 42 s | 18.2 |
| Llama 4 ScoutMeta · open weights | 132 | 0.6 s | 6.5 |
| Qwen3.5-122B-A10BAlibaba · open weights | 132 | 16 s | 15.6 |
| Gemini 3.1 Pro PreviewGoogle | 130 | 15 s | 29.7 |
| GPT-6 Luna (max reasoning)OpenAI | 127 | 116 s | 38.1 |
| GPT-6 Luna (high reasoning)OpenAI | 122 | 17 s | 32.9 |
| Qwen3-Next-80B-A3B-InstructAlibaba · open weights | 121 | 0.6 s | 9.6 |
| Muse Spark 1.3 (max reasoning)Meta | 119 | 56 s | 48.1 |
| GPT-5.3-Codex (xhigh reasoning)OpenAI | 119 | 45 s | 32.5 |
| Solar Pro 4Upstage | 119 | 18 s | 26.6 |
| GPT-6 Luna (low reasoning)OpenAI | 115 | 3.5 s | 21.5 |
| GPT-5.6 Terra (xhigh reasoning)OpenAI | 114 | 21 s | 38.0 |
| GPT-6 Luna (no reasoning)OpenAI | 113 | 0.7 s | 18.5 |
| GPT-6 Luna (xhigh reasoning)OpenAI | 112 | 38 s | 34.6 |
| Qwen3.5-9B (no reasoning)Alibaba · open weights | 110 | 0.4 s | 13.3 |
| Qwen3-Omni-30B-A3B-ThinkingAlibaba | 108 | 19 s | 7.8 |
| Qwen3-Omni-30B-A3B-InstructAlibaba | 107 | 0.9 s | 6.0 |
| GPT-5.6 Terra (high reasoning)OpenAI | 106 | 6.3 s | 34.2 |
| Claude Sonnet 5.5 (xhigh reasoning)Anthropic | 103 | 19 s | 51.9 |
| Claude Sonnet 5.5 (high reasoning)Anthropic | 100 | 9.2 s | 46.8 |
| GPT-5.6 Terra (medium reasoning)OpenAI | 100 | 1.9 s | 30.1 |
| Llama 3.3 70B InstructMeta · open weights | 98 | 0.6 s | 7.7 |
| GPT-5.6 Terra (low reasoning)OpenAI | 97 | 1.4 s | 27.5 |
| Qwen3.5-Omni-PlusAlibaba | 97 | 1.1 s | 20.4 |
| Claude Opus 5.5 (max reasoning)Anthropic | 96 | 10 min | 57.6 |
| DeepSeek-V4-Pro (0813, max reasoning)DeepSeek · open weights | 96 | 22 s | 36.0 |
| GPT-5.6 Terra (no reasoning)OpenAI | 96 | 0.7 s | 20.8 |
| Ministral 3 8BMistral AI · open weights | 93 | 0.5 s | 5.5 |
| Claude Sonnet 5.5 (medium reasoning)Anthropic | 93 | 0.9 s | 40.8 |
| DeepSeek-V4-Pro (0813, no reasoning)DeepSeek · open weights | 93 | 1.2 s | 20.4 |
| Claude Sonnet 5.5 (low reasoning)Anthropic | 93 | 0.7 s | 35.9 |
| GLM-5.3 (low reasoning)Z.ai · open weights | 87 | 25 s | 34.3 |
| Hy3Tencent · open weights | 83 | 26 s | 25.3 |
| Reka Flash 3Reka AI · open weights | 83 | 25 s | 5.6 |
| Gemma 4 26B A4B (no reasoning)Google · open weights | 82 | 0.6 s | 13.1 |
| Claude Opus 5.5 (xhigh reasoning)Anthropic | 80 | 81 s | 56.0 |
| Ministral 3 14BMistral AI · open weights | 79 | 0.5 s | 6.0 |
| GLM-5.3-FlashZ.ai · open weights | 79 | 27 s | 41.8 |
| GLM-5.3 (max reasoning)Z.ai · open weights | 78 | 28 s | 44.8 |
| Qwen3-Coder-NextAlibaba · open weights | 77 | 0.7 s | 9.2 |
| Magistral Small 1.2Mistral AI | 76 | 27 s | 8.6 |
| Claude Opus 5.5 (low reasoning)Anthropic | 76 | 6.9 s | 42.3 |
| Claude Opus 5.5 (high reasoning)Anthropic | 75 | 27 s | 53.6 |
| Grok 4.7 (xhigh reasoning)xAI | 74 | 22 s | 46.4 |
| Claude Opus 5.5 (medium reasoning)Anthropic | 74 | 14 s | 51.2 |
| Grok 4.7 (high reasoning)xAI | 74 | 18 s | 46.3 |
| Claude Fable 5.1 (max reasoning)Anthropic | 72 | 2.8 min | 53.4 |
| GPT-6 Astra (max reasoning)OpenAI | 72 | 3.7 min | 52.7 |
| Qwen3.5-9BAlibaba · open weights | 71 | 29 s | 11.2 |
| GPT-6 Astra (xhigh reasoning)OpenAI | 67 | 118 s | 52.4 |
| Granite 4.2 8BIBM · open weights | 65 | 31 s | 11.1 |
| Grok 4.7 (low reasoning)xAI | 63 | 4.0 s | 42.2 |
| Mistral Large 4Mistral AI · open weights | 60 | 34 s | 38.4 |
| Claude Fable 5.1 (xhigh reasoning)Anthropic | 60 | 47 s | 53.2 |
| Command ACohere · open weights | 59 | 0.4 s | 7.0 |
| GPT-6 Astra (high reasoning)OpenAI | 59 | 51 s | 50.9 |
| GPT-6 Astra (low reasoning)OpenAI | 56 | 1.4 s | 45.8 |
| Qwen3.7-PlusAlibaba | 56 | 37 s | 25.2 |
| Qwen3.5-397B-A17B (no reasoning)Alibaba · open weights | 56 | 1.5 s | 21.4 |
| Qwen3.8-Flash-NextAlibaba | 55 | 38 s | 39.8 |
| GPT-6 Astra (medium reasoning)OpenAI | 55 | 5.7 s | 49.6 |
| Qwen3.5-397B-A17BAlibaba · open weights | 55 | 60 s | 18.4 |
| Claude Fable 5.1 (medium reasoning)Anthropic | 54 | 6.0 s | 48.9 |
| Kimi K2.7 CodeMoonshot AI · open weights | 54 | 43 s | 25.8 |
| Gemma 4 26B A4BGoogle · open weights | 53 | 38 s | 16.7 |
| GPT-6.1 Sol (max reasoning)OpenAI | 53 | 4.2 min | 51.8 |
| Claude Fable 5.1 (low reasoning)Anthropic | 52 | 0.9 s | 46.8 |
| Claude Fable 5.1 (high reasoning)Anthropic | 51 | 10.0 s | 51.2 |
| GPT-6.1 Sol (xhigh reasoning)OpenAI | 50 | 110 s | 51.0 |
| GPT-6.1 Sol (high reasoning)OpenAI | 49 | 41 s | 50.2 |
| Gemma 4 31B (no reasoning)Google · open weights | 48 | 0.8 s | 13.9 |
| Qwen3.8-27B (low reasoning)Alibaba · open weights | 46 | 44 s | 26.2 |
| Llama 4 MaverickMeta · open weights | 46 | 0.6 s | 9.3 |
| MiMo-V2.6-FlashXiaomi · open weights | 46 | 47 s | 37.9 |
| Phi-4Microsoft · open weights | 45 | 1.0 s | 5.9 |
| GPT-6.1 Sol (low reasoning)OpenAI | 45 | 1.7 s | 42.1 |
| GPT-6.1 Sol (medium reasoning)OpenAI | 45 | 7.0 s | 47.8 |
| Qwen3.8-27B (medium reasoning)Alibaba · open weights | 43 | 48 s | 27.6 |
| Qwen3.8-27B (no reasoning)Alibaba · open weights | 43 | 1.3 s | 20.2 |
| MiMo-V2.6-ProXiaomi · open weights | 43 | 50 s | 46.3 |
| Kimi K3 (max reasoning)Moonshot AI · open weights | 41 | 52 s | 43.6 |
| Kimi K3 (low reasoning)Moonshot AI · open weights | 40 | 53 s | 30.1 |
| Qwen3.8-27B (xhigh reasoning)Alibaba · open weights | 39 | 52 s | 33.7 |
| Qwen3.8-Max (0902)Alibaba | 37 | 56 s | 45.4 |
| Qwen3.8-2.4T-A95BAlibaba · open weights | 36 | 58 s | 39.9 |
| Gemma 4 31BGoogle · open weights | 35 | 50 s | 14.7 |
| Gemma 4 E4B (no reasoning)Google | 32 | 0.4 s | 7.5 |
| Gemma 4 E4BGoogle | 28 | 72 s | 8.9 |
| Qwen3.5-4B (thinking reasoning)Alibaba | 10 | 3.4 min | 13.1 |
| Qwen3.5-4B (no reasoning)Alibaba | 9 | 0.7 s | 10.8 |
Tokens per second against intelligence
How fast each model writes for what it can do
Further up is more capable; further right writes faster. Green settings write fastest for their intelligence: none is beaten on both.
Fastest for their intelligenceOther settings (hover for names)
Each dot is one model at one reasoning setting. Green dots write fastest for their intelligence: for each, no other setting writes faster and is more capable. Tokens per second is on a log scale. It leaves out thinking time, so a high reasoning setting can write quickly and still take minutes per task.
The fastest writer at each level
Grouped by the Artificial Analysis Intelligence Index.
| Intelligence | Fastest writer | Tokens/s | Most capable | Tokens/s |
|---|---|---|---|---|
| Frontier: 50 or more | Muse Wonder (max reasoning)Meta · 52.4 | 153 | Claude Opus 5.5 (max reasoning)Anthropic · 57.6 | 96 |
| Strong: 40–50 | Muse Spark 1.3 (xhigh reasoning)Meta · 45.1 | 254 | GPT-6 Astra (medium reasoning)OpenAI · 49.6 | 55 |
| Capable: 30–40 | Beam 1.0 501B A23BReflection · 30.3 | 323 | Qwen3.8-2.4T-A95BAlibaba · 39.9 | 36 |
| Everyday: 20–30 | Nemotron 3.5 Super VLNVIDIA · 26.5 | 411 | Gemini 3.1 Pro PreviewGoogle · 29.7 | 130 |
| Light: under 20 | Trinity Large ThinkingArcee AI · 10.8 | 362 | GPT-6 Luna (no reasoning)OpenAI · 18.5 | 113 |
How to get answers faster
Time your own task
Tokens per second is only part of it. A model that thinks for longer, or writes three times as much, takes longer whatever its writing speed. Time a few real requests on two or three candidates.
Set the reasoning effort deliberately
The setting matters as much as the model: the same model can start answering in two seconds or eight minutes. Try the lowest setting that passes your task.
Stream the answer
For chat, people judge speed by when words appear. Streaming shows the answer from the first token.
Ask for less
Writing time grows with every output token. Ask for structured output, cap the length, and leave out what you won't use.
Run independent calls in parallel
Ten separate extractions sent at once finish in the time of the slowest, not of all ten in a row.
Check the host
Open-weights models run on many hosts at very different speeds; some sell a faster tier. The tokens per second here are on each model's usual API.
Quick answers
What is the fastest LLM API?
Nemotron 3.5 Super VL writes fastest, at 411 output tokens a second on its usual API, measured by Artificial Analysis.
What is the fastest capable LLM?
Muse Spark 1.3 (xhigh reasoning) writes fastest of the settings scoring 40 to 50 on the Intelligence Index, at 254 tokens a second.
How fast is the most capable model?
Claude Opus 5.5 (max reasoning) scores highest on the Intelligence Index (57.6). It writes 96 tokens a second, but thinks first: its answers typically start after 10 min.
How long does an LLM take for a real task?
On CatalogBench, turning a sparse product feed into a listing, Gemini 3.5 Flash-Lite took 3.3 s per product and Kimi K3.1 Pro 114 s, measured on our own runs.
Why isn't tokens per second the same as speed on a task?
Tokens per second is how fast a model writes once it has started. The time a task takes also depends on how long it thinks first and how many tokens it writes in all: a slower writer that solves the task in fewer tokens can finish first.
Which model is quick enough for your task?
Run the models you're considering side by side on your own prompt. See how long each answer takes, whether it passes your checks, and what it costs.
Sources and method
- Seconds and tokens per task: recorded while running our own benchmarks (CatalogBench, DeckBench, SurveyBench, BulletBench), averaged over the same tasks as their cost, including any retry needed for a usable answer. SurveyBench adds its two steps together. BulletBench is the median time per chess move. Output tokens include hidden thinking; tokenizers differ between developers, so token counts compare effort only roughly.
- Tokens per second, first token and intelligence: Artificial Analysis (data sourced from Artificial Analysis), fetched 9 Nov 2026. Medians over recent weeks on each model's usual API. Each row's Intelligence Index is for the same setting as its speed.