BenchLeader
Menu
Best LLMsChoose a modelCoding plansLeaderboardsModelsBenchmarksProvidersReleases4

How fast are LLMs? Speed statistics

Data as of 2 Nov 2026 · last updated · by BenchLeader

The median ranked model on BenchLeader generates about 83 output tokens per second and starts answering in a median 1.0 s (time to first token); measured by Artificial Analysis and OpenRouter. The fastest model BenchLeader tracks is Celeris-1 at about 1,700 tokens per second (Artificial Analysis). Among models scoring 60 or more on the BenchLeader Index, Muse Spark 1.2 is fastest at about 300 tokens per second.

Output speed vs BenchLeader Index, ranked modelsThe median ranked model on BenchLeader generates about 83 output tokens per second and starts answering in a median 1.0 s (time to first token); measured by Artificial Analysis and OpenRouter.Output speed vs BenchLeader Index, ranked models1.0101001,000303540455055606570BenchLeader IndexOutput tokens per second (log)Claude Opus 5.5: 69.2, 96Gemini 4 Carbon: 68.9, 95GPT-6 Astra: 68.2, 44GPT-6.1 Sol: 67.2, 87Claude Fable 5.1: 67.1, 70Claude Fable 5: 66.6, 54Gemini 4 Argon: 65.9, 83GPT-5.6 Sol: 65.4, 130Claude Opus 5: 65.4, 53GPT-6.1 Terra: 65.2, 110Claude Sonnet 5.5: 65.1, 140GPT-5.5: 64.4, 91Muse Wonder: 64.0, 84GPT-6 Sol: 63.5, 150GPT-5.4: 63.3, 110Claude Opus 4.7: 63.1, 49GPT-5.4 Pro: 63.1, 7.0GPT-5.3-Codex: 62.9, 110Claude Opus 4.6: 62.6, 35Gemini 3.5 Flash: 62.5, 190Gemini 3.7 Flash: 62.2, 91Muse Spark 1.1: 62.2, 210Kimi K3: 62.1, 66Gemini 3.8 Flash: 61.9, 130Muse Spark 1.3: 61.9, 190Muse Spark 1.2: 61.6, 300Kimi K3.1 Pro: 61.5, 31Gemini 3.6 Flash: 61.3, 200GPT-5.6 Terra: 61.1, 130o3: 61.0, 160Gemini 3.1 Pro: 60.6, 140Claude Opus 4.8: 60.6, 66Grok 4.6: 60.5, 70Qwen3.7: 60.3, 200GLM-5: 60.2, 62GPT-5.2: 60.1, 80GLM 5.2: 60.1, 89Grok 4.5: 59.8, 54GPT-5 Pro: 59.8, 18Claude Opus 4.5: 59.7, 57Kimi K2.6: 59.7, 41Claude Sonnet 5: 59.4, 71DeepSeek V4.1 Flash 1017: 59.2, 170Qwen3.6 Plus: 59.1, 54GLM 5.3: 59.0, 89Gemini 3 Flash: 58.9, 210Qwen3.8 Max: 58.8, 40Kimi K2.5: 58.7, 67GPT-5-Codex: 58.6, 180GPT-5.6 Luna: 58.5, 110Grok 4: 58.4, 78Grok 4.20: 58.4, 130Grok 4.20 Multi-Agent: 58.4, 360Claude Sonnet 4.6: 58.3, 44Grok 4.7: 58.2, 80GPT-5: 58.2, 110Claude Opus 4.1: 57.9, 13Claude Sonnet 4.5: 57.9, 38Qwen3.7 Plus: 57.8, 57GPT-5.1: 57.7, 140DeepSeek V4 Pro: 57.5, 42MiMo-V2.6-Pro: 57.3, 44DeepSeek V4 Flash: 57.3, 49Qwen3.5 Plus: 57.3, 71GPT-5.2-Codex: 57.3, 40Claude Haiku 5.5: 57.2, 200GPT-5.5 Instant: 57.0, 150GLM 5.3 Flash: 56.8, 76GPT-6 Luna: 56.7, 120DeepSeek V3.2: 56.7, 16Grok 4.3: 56.6, 82Qwen3.5 397B A17B: 56.6, 57MiniMax M2.7: 56.5, 54MiniMax M3: 56.4, 120GPT-5 mini: 56.4, 83Step 3.5 Flash: 56.4, 130GPT-5.4 Mini: 56.1, 300Gemma 4 26B A4B: 56.1, 64Qwen3.6 27B: 56.0, 57GPT-5 nano: 55.9, 82GLM 5.1: 55.8, 37Hy3: 55.8, 97GLM 5V Turbo: 55.8, 46DeepSeek V3.1 Terminus: 55.6, 20Hy4: 55.6, 47Nemotron 3 Ultra 550B A55B: 55.5, 110MiMo-V2.5: 55.1, 68MiMo-V2.6-Flash: 55.0, 37MiniMax-M2.1: 55.0, 55Step 5 Preview: 54.9, 150MiMo-V2.5-Pro: 54.9, 18Qwen3.8-Flash-Next: 54.7, 52Gemini 2.5 Pro: 54.5, 110Beam 1.0 501B A23B: 54.5, 190Muse Glimmer: 54.5, 110Gemini 3.1 Flash Lite: 54.3, 380Kimi K2.7 Code: 54.2, 46Qwen3.5-122B-A10B: 54.2, 77GPT-5.4 Nano: 54.1, 170Qwen3.8 27B: 54.0, 44Muse Spark Open: 53.8, 190MiniMax M2.5: 53.7, 51Qwen3 VL 235B A22B Instruct: 53.5, 40Nova 2.0 Pro Preview: 53.5, 120Kimi K2 Thinking: 53.3, 120Step 3.7 Flash: 53.0, 92Mistral Large 4: 52.8, 170MiniMax-M2: 52.6, 120Grok 3: 52.5, 78Qwen3-Next 80B A3B: 52.3, 180Inkling Small: 52.2, 230Qwen3.5-27B: 52.2, 56MiniMax M3.1 Flash: 52.1, 140Solar Pro 4: 52.0, 100Qwen3.8 2.4T A95B: 52.0, 38Qwen3 235B A22B 2507: 51.9, 68Qwen3.5 Flash: 51.9, 85Qwen3.6 35B A3B: 51.9, 130GLM 4.7: 51.8, 25Kimi K2 0905: 51.8, 51Ling 3.0 Flash: 51.8, 310GLM-4.5: 51.6, 24Gemini 3.5 Flash Lite: 51.5, 420Qwen3 32B: 51.4, 29GLM-4.6: 51.4, 35GLM 4.7 Flash: 51.2, 43Nemotron 3 Super 120B A12B: 51.2, 160Ling 3.1 Flash: 51.0, 220GLM 4.5V: 50.8, 45o3-mini: 50.7, 230Qwen3.5-9B: 50.5, 63Inkling: 50.3, 210Qwen3.5-35B-A3B: 50.1, 25gpt-oss-120b: 49.8, 220Gemma 4 31B: 49.8, 62MiniMax-M1: 49.7, 120K2 Horizon 375B A23B: 49.7, 720DeepSeek R1 0528: 49.4, 25Qwen3.5-4B: 49.4, 18Mercury 2: 49.2, 360Seed-OSS-36B-Instruct: 49.2, 39Qwen3 Coder 480B: 48.8, 69Nova 2 Lite: 48.8, 110GPT-4.1: 48.5, 180Qwen-Plus: 48.4, 50Mistral Medium 3.1: 48.3, 100Qwen3 30B A3B 2507: 48.1, 200Claude Haiku 4.5: 47.6, 81Nemotron 3 Nano 30B A3B: 47.5, 240Mistral Medium 3.5: 47.3, 180Ring-2.6-1T: 47.0, 130Ling Flash 2.0: 46.9, 7.1Qwen3 14B: 46.8, 60GLM-4.5-Air: 46.7, 82North Mini Code: 46.6, 87GPT-4.1 mini: 46.3, 170Command A+: 45.9, 150Nemotron 3.5 Super VL: 45.9, 480Mistral Large 3: 45.7, 94Beam 1.0: 45.7, 360gpt-oss-20b: 45.7, 250Mistral Medium 3: 44.4, 41Qwen3.6 Flash: 44.0, 64GPT-4o: 43.3, 180Qwen2.5 72B: 43.2, 27Devstral 2: 42.8, 100Qwen2.5 Coder 32B Instruct: 42.7, 31Qwen3 Coder Next: 42.4, 80Llama 4 Maverick: 42.2, 60Command A: 42.2, 15Mistral Small 3.2: 42.1, 40Nemotron 3.5 Lightning: 41.4, 290GPT-4 Turbo: 41.3, 120Granite 4.2 8B: 41.1, 61Nova Pro: 40.9, 130Llama 3.1 70B: 40.9, 41Mistral Large 2: 40.9, 49Gemma 2 27B: 40.9, 47Solar Pro 3: 40.8, 140Mistral Large 24.02: 40.7, 60Mistral Small 4: 40.4, 160Llama 3.3 70B: 40.0, 91Magistral Small 1.2: 40.0, 74Gemma 3 27B: 39.9, 23Ministral 3 14B: 39.9, 61Qwen2.5 7B Instruct: 39.9, 78Phi-4: 39.5, 41Mistral Small 3: 39.4, 49Nova Micro: 39.2, 270Ministral 3 8B: 38.7, 91Llama 4 Scout: 38.6, 110Mistral Small 3.1: 38.5, 38Nova Lite: 38.3, 170Gemma 3 12B: 37.9, 30Mixtral 8x22B Instruct: 37.5, 110GPT-4.1 nano: 37.0, 280Ministral 3 3B: 36.8, 180Llama 3.2 3B Instruct: 35.8, 120Llama 3.2 Instruct 11B (Vision): 35.7, 25GPT-4o mini: 35.4, 110Gemma 3 4B: 34.9, 61Llama 3.1 8B: 33.7, 140Mixtral 8x7B: 30.9, 25Llama 3.2 1B Instruct: 30.8, 97K2 Horizon 375B A23BGemini 3.5 Flash LiteGrok 4.20 Multi-AgentMuse Spark 1.2Claude Sonnet 5.5Claude Opus 5.5Speeds measured by Artificial Analysis and OpenRouterIndex: BenchLeader v1.4 · data as of 2 Nov 2026www.benchleader.com/stats/llm-speed
The median ranked model on BenchLeader generates about 83 output tokens per second and starts answering in a median 1.0 s (time to first token); measured by Artificial Analysis and OpenRouter. Each dot is a ranked model with a measured speed; named dots are the fastest in each Index tier. Chart image (SVG)

Key statistics

Each statistic has its own link (#) to cite it directly.

  • The median ranked model generates about 83 output tokens per second, across 204 ranked models with a measured speed (measured by Artificial Analysis and OpenRouter; BenchLeader, 2 Nov 2026). #
  • The median time to first token is 1.0 s, across 187 ranked models (measured by OpenRouter and Artificial Analysis; BenchLeader, 2 Nov 2026). #
  • The fastest model BenchLeader tracks is Celeris-1 by Celeris, at about 1,700 tokens per second (Artificial Analysis; BenchLeader, 2 Nov 2026). #
  • The fastest ranked model scoring 60 or more on the BenchLeader Index is Muse Spark 1.2, at about 300 tokens per second (Artificial Analysis; BenchLeader, 2 Nov 2026). #
  • The lowest time to first token among ranked models scoring 60 or more is GLM 5.2's 0.73 s (OpenRouter; BenchLeader, 2 Nov 2026). #
  • The lowest time to first token of any ranked model is Qwen3.5-9B's 0.18 s (OpenRouter; BenchLeader, 2 Nov 2026). #
  • Claude Opus 5.5, the BenchLeader Index leader, generates about 96 tokens per second, measured by Artificial Analysis, with a 3.4 s time to first token, measured by OpenRouter (BenchLeader, 2 Nov 2026). #

Fastest model at each quality tier

Index tierModels with a speedFastest modelTokens per secondMeasured by
65+11Claude Sonnet 5.5 Anthropic140Artificial Analysis
60–6526Muse Spark 1.2 Meta300Artificial Analysis
55–6052Grok 4.20 Multi-Agent SpaceXAI360OpenRouter
50–5544Gemini 3.5 Flash Lite Google420Artificial Analysis
45–5028K2 Horizon 375B A23B MBZUAI720Artificial Analysis

Fastest models by output speed

Ranked models only (those with a BenchLeader Index). The speed leaderboard also lists models without an Index, which is why the fastest model overall may not appear here.

#ModelIndexTokens per secondMeasured by
1K2 Horizon 375B A23B MBZUAI49.7720Artificial Analysis
2Nemotron 3.5 Super VL NVIDIA45.9480Artificial Analysis
3Gemini 3.5 Flash Lite Google51.5420Artificial Analysis
4Gemini 3.1 Flash Lite Google54.3380Artificial Analysis
5Grok 4.20 Multi-Agent SpaceXAI58.4360OpenRouter
6Beam 1.0 Reflection45.7360Artificial Analysis
7Mercury 2 Inception49.2360OpenRouter
8Ling 3.0 Flash Ant Group51.8310Artificial Analysis
9GPT-5.4 Mini OpenAI56.1300Artificial Analysis
10Muse Spark 1.2 Meta61.6300Artificial Analysis

Quickest to the first token

Ranked models with the lowest time to first token; see the latency leaderboard for all of them.

#ModelIndexTime to first tokenMeasured by
1Qwen3.5-9B Alibaba50.50.18 sOpenRouter
2Llama 3.2 1B Instruct Meta30.80.23 sOpenRouter
3Qwen2.5 7B Instruct Alibaba39.90.25 sOpenRouter
4Nemotron 3 Ultra 550B A55B NVIDIA55.50.27 sOpenRouter
5Command A+ Cohere45.90.28 sOpenRouter
6Nemotron 3.5 Lightning NVIDIA41.40.29 sOpenRouter
7Qwen3 30B A3B 2507 Alibaba48.10.30 sOpenRouter
8Granite 4.2 8B IBM41.10.33 sOpenRouter
9Ministral 3 3B Mistral AI36.80.34 sOpenRouter
10Nova Micro Amazon39.20.35 sOpenRouter

How these numbers are made

Output speed is the median number of tokens a model generates per second once it starts answering: Artificial Analysis' measured median where it covers a model, otherwise OpenRouter's live-traffic median smoothed over seven days. Time to first token is credited to its measurer in each table; OpenRouter's figure leaves out any thinking phase. Medians are over ranked models (those with a BenchLeader Index), at each model's best-scoring configuration.

Speed depends on the provider serving the model, the load and the prompt, so treat a figure as typical rather than guaranteed. Full sourcing is in the methodology.

Cite these statistics

Speeds are measured by Artificial Analysis and OpenRouter and shown here with attribution; they stay under the measurers' terms and are not offered for reuse. BenchLeader's own figures on this page are the Index scores and the medians. Cite the page below.

Citation

BenchLeader, “LLM speed statistics”, data as of 2 Nov 2026, https://www.benchleader.com/stats/llm-speed.

Questions people ask

How fast are LLMs in tokens per second?
The median ranked model on BenchLeader generates about 83 output tokens per second (measured by Artificial Analysis and OpenRouter). The fastest tracked model, Celeris-1, reaches about 1,700.
What is a typical time to first token?
A median 1.0 s across 187 ranked models (measured by OpenRouter and Artificial Analysis). OpenRouter's first-token figure leaves out any thinking phase.
What is the fastest high-quality LLM?
Muse Spark 1.2 by Meta: about 300 tokens per second (Artificial Analysis) among ranked models scoring 60 or more on the BenchLeader Index.
Where do these speeds come from?
Output speed is Artificial Analysis' measured median where it covers a model, otherwise OpenRouter's live-traffic median smoothed over seven days; time to first token is mostly OpenRouter's. Each figure is credited to its measurer, whose terms apply; BenchLeader does not relicense them.

Related: speed leaderboard, latency leaderboard, fastest LLM, LLM price statistics, all statistics.