API providers
Who actually serves the models, what they charge for them, and how fast they answer. Prices and endpoint telemetry are ingested from OpenRouter — we do not run our own provider benchmarks.
77 hosts serving tracked models
All 77 of 77 providers
| Provider | Models served | Mean output t/s | Mean TTFT | Lowest blended $/M |
|---|---|---|---|---|
OpenAI
· first party |
101 | 53 | 4.0s | $0.069 |
DeepInfra
|
80 | 45 | 1.0s | $0.022 |
Novita
|
79 | 39 | 1.7s | $0.000 |
Google
· first party |
59 | 83 | 2.0s | $0.087 |
Alibaba
· first party |
56 | 54 | 1.0s | $0.055 |
Azure
· first party |
51 | 46 | 5.1s | $0.14 |
| Pa Parasail | 45 | 47 | 0.9s | $0.030 |
Together
|
42 | 61 | 1.0s | $0.075 |
| Si SiliconFlow | 41 | 33 | 2.1s | $0.075 |
| Ve Venice | 36 | 40 | 1.3s | $0.11 |
| AB Amazon Bedrock · first party | 32 | 64 | 2.2s | $0.061 |
| At AtlasCloud | 32 | 45 | 2.0s | $0.000 |
| Ph Phala | 28 | 41 | 2.4s | $0.068 |
| St StreamLake | 28 | 38 | 2.0s | $0.084 |
Anthropic
· first party |
27 | 68 | 2.7s | $1.00 |
| Co CoreWeave | 27 | 75 | 0.6s | $0.055 |
Google AI Studio
|
25 | 115 | 3.0s | $0.000 |
Mistral
· first party |
25 | 61 | 1.1s | $0.10 |
| Gm GMICloud | 24 | 33 | 3.4s | $0.000 |
Fireworks
|
23 | 66 | 1.5s | $0.12 |
| Di DigitalOcean | 22 | 22 | 1.3s | $0.093 |
| Cl Cloudflare | 19 | 38 | 2.9s | $0.041 |
| Ne Nebius | 17 | 48 | 2.4s | $0.10 |
| Ne NextBit | 17 | 41 | 1.3s | $0.060 |
BaseTen
|
13 | 77 | 1.0s | $0.16 |
| Mo Morph | 13 | 511 | 2.5s | $0.17 |
Z.AI
· first party |
13 | 23 | 8.0s | $0.12 |
Baidu
· first party |
11 | 63 | 1.0s | $0.11 |
| Ch Chutes | 11 | 26 | 2.7s | $0.18 |
| Ak AkashML | 10 | 45 | 1.5s | $0.040 |
| Cr Crusoe | 10 | 66 | 0.7s | $0.087 |
| IN Io Net | 10 | 36 | 1.9s | $0.073 |
| Da Darkbloom | 9 | 23 | 2.8s | $0.000 |
Groq
|
9 | 194 | 0.4s | $0.058 |
| CP Claude Platform on AWS | 9 | 43 | 3.0s | $4.00 |
| Fr Friendli | 8 | 69 | 2.5s | $0.20 |
| Io Ionstream | 8 | 43 | 1.0s | $0.17 |
| M2 Mancer 2 | 8 | 32 | 0.9s | $0.17 |
Minimax
· first party |
8 | 46 | 1.1s | $0.42 |
Nvidia
· first party |
8 | 40 | 11.0s | $0.000 |
| SR Sail Research | 8 | 35 | 1.3s | $0.094 |
| Re Reka · first party | 7 | 43 | 1.7s | $0.10 |
SambaNova
|
7 | 87 | 1.3s | $0.34 |
| Wa Wafer | 7 | 46 | 1.7s | $0.14 |
xAI
· first party |
7 | 116 | 5.3s | $1.25 |
| In Inceptron | 6 | 44 | 0.8s | $0.17 |
| Po Poolside | 6 | 50 | 1.5s | $0.000 |
| Re Relace | 6 | 1,244 | 1.1s | $0.094 |
| Se Seed · first party | 6 | 52 | 1.0s | $0.13 |
| Ma Makora | 6 | 94 | 1.1s | $0.12 |
Cohere
· first party |
5 | 48 | 0.3s | $0.000 |
| De Decart | 5 | 119 | 1.9s | $0.000 |
DeepSeek
· first party |
5 | 62 | 1.0s | $0.33 |
| Ma Mara | 5 | 107 | 2.8s | $0.30 |
Meta
· first party |
5 | 105 | 3.9s | $0.12 |
| Mo ModelRun | 5 | 106 | 0.7s | $0.81 |
Perplexity
· first party |
5 | 31 | 27.0s | $1.00 |
Tencent
· first party |
5 | 21 | 1.5s | $0.077 |
| Ai AionLabs | 4 | 45 | 1.0s | $0.88 |
| Am Ambient | 4 | 28 | 3.2s | $0.10 |
Moonshot AI
· first party |
4 | 27 | 2.3s | $1.20 |
| NA Nex AGI | 4 | 68 | 2.7s | $0.000 |
| Mo Modal | 4 | 111 | 0.8s | $0.24 |
Cerebras
|
3 | 476 | 0.4s | $0.45 |
| In Inception | 3 | 255 | 1.1s | $0.068 |
| Op OpenInference | 3 | 24 | 2.7s | $0.077 |
| In Inflection | 2 | — | — | $4.38 |
| SA Sakana AI | 2 | 11 | 3.2s | $1.71 |
Upstage
· first party |
2 | 30 | 1.9s | $0.052 |
| Xi Xiaomi · first party | 2 | 28 | 4.0s | $0.17 |
| TM Thinking Machines · first party | 2 | 72 | 1.3s | $0.000 |
| Ai AI21 · first party | 1 | — | — | $3.50 |
| AA Arcee AI | 1 | 101 | 0.1s | $0.39 |
| Pe Perceptron | 1 | 30 | 0.6s | $0.49 |
| St StepFun | 1 | 62 | 3.7s | $0.44 |
| Li Liquid | 1 | 109 | 0.7s | $0.000 |
| St Stealth | 1 | 23 | 5.7s | $0.000 |
How to read this. A host appears here once we have seen it serve at least one tracked model.
Mean output t/s and mean TTFT average that host's measurements across every model it serves, so a host
that only serves small models will look faster than one carrying frontier models — compare hosts on a single model
page for a like-for-like reading. Lowest blended $/M is the cheapest blended rate that host publishes for any
model, at a 3:1 input:output blend. Figures are ingested from OpenRouter endpoint listings and telemetry, with
metadata enriched from models.dev. See methodology and
attribution.