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
- 128
- model and reasoning setting
- Speeds measured
- 14 Oct 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 (22 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 | Reliably publish-ready |
|---|---|---|---|
| 1 | Gemini 3.5 Flash-Lite | 3.3 s | 17.9% |
| 2 | Mistral Medium 3.5 | 4.9 s | 17.9% |
| 3 | Gemini 4 Argon | 7.3 s | 60.7% |
| 4 | Claude Haiku 4.5 | 8.1 s | 1.8% |
| 5 | GPT-6 Luna | 8.2 s | 55.4% |
| 6 | Mistral Large 4 | 8.2 s | 7.1% |
| 7 | Claude Sonnet 5.5 | 9.9 s | 32.1% |
| 8 | GPT-6.1 Sol | 10 s | 75.0% |
| 9 | Muse Wonder | 10 s | 66.1% |
| 10 | GPT-6 Sol | 11 s | 66.1% |
| 11 | Claude Haiku 5.5 | 11 s | 26.8% |
| 12 | GPT-6 Astra | 12 s | 75.0% |
| 13 | Claude Opus 5.5 | 17 s | 55.4% |
| 14 | DeepSeek-V4.1-Flash | 18 s | 41.1% |
| 15 | Gemini 3.8 Flash | 19 s | 35.7% |
| 16 | Claude Fable 5.1 | 22 s | 26.8% |
| 17 | Gemini 3.1 Pro Preview | 22 s | 44.6% |
| 18 | Grok 4.7 | 42 s | 57.1% |
| 19 | Muse Spark 1.3 | 48 s | 48.2% |
| 20 | GLM-5V-Turbo | 62 s | 1.8% |
| 21 | Kimi K3 | 67 s | 41.1% |
| 22 | Qwen3.8-Max (0902) | 78 s | 28.6% |
DeckBench: turning an analysis into a deck (21 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 | Deck rating |
|---|---|---|---|
| 1 | Gemini 3.5 Flash-Lite | 15 s | 654 |
| 2 | Mistral Medium 3.5 | 25 s | 629 |
| 3 | Muse Wonder | 29 s | 1,297 |
| 4 | Claude Haiku 4.5 | 52 s | 678 |
| 5 | Claude Haiku 5.5 | 62 s | 1,166 |
| 6 | Gemini 4 Argon | 68 s | 998 |
| 7 | Gemini 3.1 Pro Preview | 78 s | 763 |
| 8 | Claude Sonnet 5.5 | 83 s | 1,379 |
| 9 | Muse Spark 1.3 | 89 s | 852 |
| 10 | Gemini 3.8 Flash | 93 s | 805 |
| 11 | Claude Opus 5.5 | 98 s | 1,409 |
| 12 | GPT-6 Luna | 101 s | 1,331 |
| 13 | Mistral Large 4 | 114 s | 578 |
| 14 | Claude Fable 5.1 | 2.6 min | 1,351 |
| 15 | GPT-6 Sol | 2.8 min | 1,500 |
| 16 | GPT-6.1 Sol | 3.0 min | 1,430 |
| 17 | GPT-6 Astra | 5.2 min | 1,553 |
| 18 | DeepSeek-V4-Pro (0423) | 5.5 min | 564 |
| 19 | Grok 4.7 | 6.2 min | 1,290 |
| 20 | Kimi K3 | 7.3 min | 1,231 |
| 21 | Qwen3.8-Max (0902) | 9.1 min | 1,053 |
SurveyBench: planning a user survey and analysing its responses (10 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) | Analysis rating |
|---|---|---|---|
| 1 | Gemini 3.5 Flash-Lite | 16 s | 623 |
| 2 | Mistral Medium 3.5 | 48 s | 544 |
| 3 | Gemini 4 Argon | 74 s | 964 |
| 4 | Muse Wonder | 83 s | 1,494 |
| 5 | Claude Haiku 4.5 | 93 s | 504 |
| 6 | Claude Haiku 5.5 | 2.5 min | 1,276 |
| 7 | Gemini 3.1 Pro Preview | 2.8 min | 776 |
| 8 | Claude Opus 5.5 | 3.0 min | 1,496 |
| 9 | GPT-6.1 Sol | 3.7 min | 1,450 |
| 10 | Mistral Large 4 | 7.3 min | 783 |
BulletBench: choosing a chess move on the Lightning 10+1 clock (23 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
| # | Model and setting | Seconds per move (median) | Ladder Elo |
|---|---|---|---|
| 1 | Seed-2.0-Mini (minimal reasoning) | 0.3 s | 571 |
| 2 | Mercury Decide | 0.4 s | 614 |
| 3 | Mistral Medium 3.5 | 0.5 s | 529 |
| 4 | Jev 1.13 | 0.5 s | 571 |
| 5 | GPT-6 Luna Decisions | 0.5 s | 529 |
| 6 | Clef Flash | 0.5 s | 529 |
| 7 | Mistral Small 4 (no reasoning) | 0.6 s | 571 |
| 8 | Inkling Small (no reasoning) | 0.6 s | 571 |
| 9 | Decider V1 27B | 0.6 s | 420 |
| 10 | d1 | 0.6 s | 572 |
| 11 | Clef | 0.7 s | 529 |
| 12 | Gemini 3.1 Flash-Lite (minimal reasoning) | 0.8 s | 780 |
| 13 | Mistral Large 4 | 0.8 s | 249 |
| 14 | Gemini 3.5 Flash-Lite | 0.9 s | 892 |
| 15 | GPT-5.4 mini (no reasoning) | 0.9 s | 571 |
| 16 | Kev 4B | 1.0 s | 33 |
| 17 | GPT-5.4 nano (no reasoning) | 1.0 s | 197 |
| 18 | Gemini 3.5 Flash (minimal reasoning) | 1.2 s | 727 |
| 19 | GPT-6.1 Sol Ultrafast (low reasoning) | 1.2 s | 584 |
| 20 | Claude Haiku 4.5 | 1.3 s | 33 |
| 21 | Gemini 3.6 Flash (minimal reasoning) | 1.5 s | 33 |
| 22 | Gemini 3.8 Flash (low reasoning) | 1.8 s | 197 |
| 23 | Claude Sonnet 5.5 (low reasoning) | 2.1 s | 33 |
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; they appear here as models are benchmarked 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 |
|---|---|---|---|
| Gemini 3.5 Flash-LiteGoogle | 393 | 5.0 s | 22.2 |
| Trinity Large ThinkingArcee AI · open weights | 367 | 6.6 s | 10.8 |
| gpt-oss-20b (low reasoning)OpenAI · open weights | 292 | 7.3 s | 10.0 |
| Nemotron 3 Nano 30B A3B (no reasoning)NVIDIA · open weights | 261 | 0.3 s | 6.8 |
| Nova Micro 1.0Amazon | 257 | 0.6 s | 5.9 |
| gpt-oss-20b (high reasoning)OpenAI · open weights | 250 | 8.4 s | 9.0 |
| gpt-oss-120b (low reasoning)OpenAI · open weights | 242 | 8.8 s | 10.2 |
| Ministral 3 3BMistral AI · open weights | 241 | 0.4 s | 4.8 |
| Claude Haiku 5.5 (max reasoning)Anthropic | 239 | 3.9 min | 43.4 |
| Nemotron 3 Nano 30B A3B (thinking reasoning)NVIDIA · open weights | 229 | 9.2 s | 8.9 |
| Qwen3.5-Omni-FlashAlibaba | 225 | 0.9 s | 12.5 |
| Nova 2 Lite (no reasoning)Amazon | 222 | 0.8 s | 8.7 |
| gpt-oss-120b (high reasoning)OpenAI · open weights | 221 | 9.6 s | 11.6 |
| Inkling (xhigh reasoning)Thinking Machines · open weights | 210 | 12 s | 25.0 |
| Muse Spark 1.3 (xhigh reasoning)Meta | 206 | 29 s | 45.1 |
| Nova 2 Lite (low reasoning)Amazon | 203 | 14 s | 11.8 |
| Nova 2 Lite (medium reasoning)Amazon | 199 | 20 s | 12.5 |
| Nova 2 Lite (high reasoning)Amazon | 199 | 20 s | 13.4 |
| Inkling SmallThinking Machines · open weights | 186 | 13 s | 25.7 |
| DeepSeek-V4.1-Flash (max reasoning)DeepSeek · open weights | 184 | 12 s | 39.5 |
| Mistral Small 4Mistral AI · open weights | 182 | 12 s | 11.3 |
| DeepSeek-V4-Flash-Vision-Exp (max reasoning)DeepSeek · open weights | 173 | 12 s | 34.8 |
| Mistral Small 4 (no reasoning)Mistral AI · open weights | 169 | 0.7 s | 9.0 |
| Mistral Medium 3.5Mistral AI | 165 | 13 s | 14.2 |
| Claude Haiku 5.5 (xhigh reasoning)Anthropic | 164 | 41 s | 41.2 |
| DeepSeek-V4.1-Flash (no reasoning)DeepSeek · open weights | 162 | 0.9 s | 24.7 |
| GPT-5.5 Instant (2026-06-26)OpenAI | 158 | 13 s | 26.0 |
| Gemma 4 12B (no reasoning)Google | 154 | 1.4 s | 9.4 |
| Claude Haiku 5.5 (high reasoning)Anthropic | 153 | 14 s | 37.8 |
| Nemotron 3 Super 120B A12BNVIDIA · open weights | 152 | 14 s | 12.8 |
| Claude Haiku 5.5 (medium reasoning)Anthropic | 148 | 6.7 s | 34.5 |
| Gemma 4 12BGoogle | 147 | 15 s | 14.2 |
| Qwen3.5-122B-A10B (no reasoning)Alibaba · open weights | 147 | 1.1 s | 17.7 |
| Claude Haiku 5.5 (low reasoning)Anthropic | 145 | 3.8 s | 29.5 |
| Qwen3.6-35B-A3B (no reasoning)Alibaba · open weights | 142 | 1.1 s | 15.2 |
| Muse Glimmer (high reasoning)Meta | 140 | 15 s | 17.5 |
| Gemini 3.8 Flash (high reasoning)Google | 138 | 14 s | 40.9 |
| Gemini 3.1 Pro PreviewGoogle | 135 | 17 s | 29.7 |
| Qwen3.6-35B-A3BAlibaba · open weights | 135 | 41 s | 18.2 |
| Claude Sonnet 5.5 (max reasoning)Anthropic | 133 | 4.5 min | 56.0 |
| Qwen3.5-122B-A10BAlibaba · open weights | 132 | 16 s | 15.6 |
| GPT-6 Luna (max reasoning)OpenAI | 128 | 86 s | 38.1 |
| GPT-6 Luna (xhigh reasoning)OpenAI | 123 | 18 s | 34.6 |
| GPT-6 Luna (no reasoning)OpenAI | 122 | 0.9 s | 18.5 |
| GPT-6 Luna (high reasoning)OpenAI | 120 | 8.6 s | 32.9 |
| o3OpenAI | 120 | 5.5 s | 20.2 |
| Muse Wonder (max reasoning)Meta | 117 | 48 s | 52.4 |
| Ministral 3 8BMistral AI · open weights | 114 | 0.5 s | 5.5 |
| GPT-6 Luna (low reasoning)OpenAI | 112 | 2.5 s | 21.5 |
| Muse Spark 1.3 (max reasoning)Meta | 112 | 53 s | 48.1 |
| Qwen3-Omni-30B-A3B-ThinkingAlibaba | 111 | 19 s | 7.8 |
| Qwen3-Omni-30B-A3B-InstructAlibaba | 108 | 0.9 s | 6.0 |
| GPT-5.6 Terra (max reasoning)OpenAI | 106 | 70 s | 42.1 |
| Gemma 4 26B A4B (no reasoning)Google · open weights | 101 | 0.6 s | 13.1 |
| MiniMax-M3MiniMax · open weights | 101 | 21 s | 29.2 |
| Llama 4 ScoutMeta · open weights | 100 | 0.7 s | 8.1 |
| Claude Opus 5.5 (max reasoning)Anthropic | 96 | 8.0 min | 57.6 |
| Llama 3.3 70B InstructMeta · open weights | 95 | 0.7 s | 7.7 |
| DeepSeek-V4-Pro (0813, max reasoning)DeepSeek · open weights | 95 | 22 s | 36.0 |
| Reka Flash 3Reka AI · open weights | 94 | 22 s | 5.6 |
| Claude Sonnet 5.5 (xhigh reasoning)Anthropic | 94 | 17 s | 51.9 |
| GPT-5.6 Terra (high reasoning)OpenAI | 94 | 1.8 s | 34.2 |
| GPT-5.6 Terra (xhigh reasoning)OpenAI | 93 | 7.8 s | 38.0 |
| GPT-5.6 Terra (medium reasoning)OpenAI | 92 | 1.5 s | 30.1 |
| GPT-5.3-Codex (xhigh reasoning)OpenAI | 92 | 27 s | 32.5 |
| GPT-5.6 Terra (no reasoning)OpenAI | 92 | 0.8 s | 20.8 |
| GPT-5.6 Terra (low reasoning)OpenAI | 92 | 1.4 s | 27.5 |
| Hy3Tencent · open weights | 90 | 24 s | 25.3 |
| Claude Sonnet 5.5 (high reasoning)Anthropic | 90 | 4.9 s | 46.8 |
| DeepSeek-V4-Pro (0813, no reasoning)DeepSeek · open weights | 89 | 1.1 s | 20.4 |
| Claude Sonnet 5.5 (medium reasoning)Anthropic | 89 | 0.6 s | 40.8 |
| Claude Sonnet 5.5 (low reasoning)Anthropic | 89 | 0.6 s | 35.9 |
| Solar Pro 4Upstage | 88 | 24 s | 28.2 |
| Qwen3.5-Omni-PlusAlibaba | 86 | 1.2 s | 20.4 |
| Magistral Small 1.2Mistral AI | 79 | 26 s | 8.6 |
| GLM-5.3 (low reasoning)Z.ai · open weights | 78 | 29 s | 34.3 |
| Qwen3-Coder-NextAlibaba · open weights | 77 | 0.8 s | 9.2 |
| Claude Opus 5.5 (xhigh reasoning)Anthropic | 75 | 58 s | 56.0 |
| Claude Opus 5.5 (low reasoning)Anthropic | 70 | 2.8 s | 42.3 |
| Claude Opus 5.5 (high reasoning)Anthropic | 69 | 17 s | 53.6 |
| GLM-5.3 (max reasoning)Z.ai · open weights | 69 | 32 s | 44.8 |
| Claude Opus 5.5 (medium reasoning)Anthropic | 69 | 8.6 s | 51.2 |
| Command ACohere · open weights | 68 | 0.4 s | 7.0 |
| Grok 4.7 (high reasoning)xAI | 65 | 9.1 s | 46.3 |
| Claude Fable 5.1 (max reasoning)Anthropic | 65 | 2.8 min | 53.4 |
| Gemma 4 26B A4BGoogle · open weights | 65 | 32 s | 16.7 |
| Ministral 3 14BMistral AI · open weights | 63 | 0.8 s | 6.0 |
| Grok 4.7 (xhigh reasoning)xAI | 63 | 15 s | 46.4 |
| Granite 4.2 8BIBM · open weights | 62 | 32 s | 11.1 |
| GLM-5.3-FlashZ.ai · open weights | 61 | 35 s | 41.8 |
| Grok 4.7 (low reasoning)xAI | 60 | 2.6 s | 42.2 |
| Kimi K2.7 CodeMoonshot AI · open weights | 59 | 39 s | 25.8 |
| Qwen3.5-397B-A17B (no reasoning)Alibaba · open weights | 57 | 1.6 s | 21.4 |
| Qwen3.5-397B-A17BAlibaba · open weights | 56 | 58 s | 18.4 |
| GPT-6.1 Sol (max reasoning)OpenAI | 56 | 2.8 min | 51.8 |
| Qwen3.7-PlusAlibaba | 55 | 38 s | 25.2 |
| Qwen3.8-Flash-NextAlibaba | 53 | 39 s | 39.8 |
| GPT-6 Astra (max reasoning)OpenAI | 53 | 2.9 min | 52.7 |
| Llama 4 MaverickMeta · open weights | 53 | 0.7 s | 10.0 |
| GPT-6.1 Sol (xhigh reasoning)OpenAI | 52 | 66 s | 51.0 |
| Claude Fable 5.1 (xhigh reasoning)Anthropic | 52 | 37 s | 53.2 |
| GPT-6.1 Sol (medium reasoning)OpenAI | 51 | 4.2 s | 47.8 |
| Claude Fable 5.1 (medium reasoning)Anthropic | 51 | 2.1 s | 48.9 |
| MiMo-V2.6-FlashXiaomi · open weights | 51 | 42 s | 37.9 |
| GPT-6.1 Sol (high reasoning)OpenAI | 50 | 18 s | 50.2 |
| GPT-6 Astra (xhigh reasoning)OpenAI | 50 | 72 s | 52.4 |
| Qwen3.8-27B (no reasoning)Alibaba · open weights | 50 | 1.2 s | 20.2 |
| GPT-6.1 Sol (low reasoning)OpenAI | 49 | 1.4 s | 42.1 |
| GPT-6 Astra (high reasoning)OpenAI | 49 | 22 s | 50.9 |
| Claude Fable 5.1 (low reasoning)Anthropic | 48 | 0.5 s | 46.8 |
| Gemma 4 31B (no reasoning)Google · open weights | 48 | 1.0 s | 13.9 |
| Claude Fable 5.1 (high reasoning)Anthropic | 47 | 2.4 s | 51.2 |
| Qwen3.8-27B (low reasoning)Alibaba · open weights | 47 | 44 s | 26.2 |
| GPT-6 Astra (medium reasoning)OpenAI | 47 | 4.8 s | 49.6 |
| Kimi K3 (max reasoning)Moonshot AI · open weights | 47 | 45 s | 43.6 |
| Qwen3.8-27B (xhigh reasoning)Alibaba · open weights | 47 | 44 s | 33.7 |
| Qwen3.8-27B (medium reasoning)Alibaba · open weights | 46 | 45 s | 27.6 |
| GPT-6 Astra (low reasoning)OpenAI | 45 | 1.5 s | 45.8 |
| Phi-4Microsoft · open weights | 45 | 1.0 s | 5.9 |
| Kimi K3 (low reasoning)Moonshot AI · open weights | 42 | 50 s | 30.1 |
| Qwen3.8-2.4T-A95BAlibaba · open weights | 41 | 51 s | 39.9 |
| Qwen3.8-Max (0902)Alibaba | 40 | 52 s | 45.4 |
| Gemma 4 E4B (no reasoning)Google | 39 | 0.4 s | 7.5 |
| MiMo-V2.6-ProXiaomi · open weights | 38 | 55 s | 46.3 |
| Gemma 4 31BGoogle · open weights | 36 | 50 s | 14.7 |
| Gemma 4 E4BGoogle | 35 | 58 s | 8.9 |
| Qwen3.5-4B (thinking reasoning)Alibaba | 29 | 71 s | 13.1 |
| Qwen3.5-4B (no reasoning)Alibaba | 19 | 0.6 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 | Claude Sonnet 5.5 (max reasoning)Anthropic · 56.0 | 133 | Claude Opus 5.5 (max reasoning)Anthropic · 57.6 | 96 |
| Strong: 40–50 | Claude Haiku 5.5 (max reasoning)Anthropic · 43.4 | 239 | GPT-6 Astra (medium reasoning)OpenAI · 49.6 | 47 |
| Capable: 30–40 | DeepSeek-V4.1-Flash (max reasoning)DeepSeek · 39.5 | 184 | Qwen3.8-2.4T-A95BAlibaba · 39.9 | 41 |
| Everyday: 20–30 | Gemini 3.5 Flash-LiteGoogle · 22.2 | 393 | Gemini 3.1 Pro PreviewGoogle · 29.7 | 135 |
| Light: under 20 | Trinity Large ThinkingArcee AI · 10.8 | 367 | GPT-6 Luna (no reasoning)OpenAI · 18.5 | 122 |
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?
Gemini 3.5 Flash-Lite writes fastest, at 393 output tokens a second on its usual API, measured by Artificial Analysis.
What is the fastest capable LLM?
Claude Haiku 5.5 (max reasoning) writes fastest of the settings scoring 40 to 50 on the Intelligence Index, at 239 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 8.0 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 Qwen3.8-Max (0902) 78 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, and see how long each answer takes. Or ask us to find the fastest model that's good enough on your real data.
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 14 Oct 2026. Medians over recent weeks on each model's usual API. Each row's Intelligence Index is for the same setting as its speed.