Hayz Labs · https://hayzlabs.com/reviews/top/model-hosts/
reviews · top 20 list · checked 18 Oct
Top 19 places to run AI models
Services that run AI models for you through an API, besides the model makers themselves, often faster or cheaper.
Best: DeepInfra, Crusoe Managed Inference and 1 more (tied, scoring 90). 16 of 19 scored.
Highest score first; on a tie, the better price for the same model. Open the Table to see the number that placed each one and how it was scored.
- 1DeepInfraDeepInfra
Among the cheapest places to run popular open models, at above-median speed
Best for Cheap, high-volume use of popular open models like DeepSeek, Qwen and GLMYou give up No free tier; you need a card or prepay before your first call
90ChecklistPay as you go - 2Crusoe Managed InferenceCrusoe
Low-priced open models with quick first responses; also rents GPUs
Best for Chat apps on open models where a fast first reply mattersYou give up Costs a bit more per token than DeepInfra at the same score
90ChecklistPay as you go - 3Azure AI FoundryMicrosoft
For teams already on Azure who want GPT, Claude and open models under one contract
Best for Companies already buying from Microsoft that need big-name models with fast outputYou give up Slowest first reply on this list, though output is fast once it starts
90ChecklistNo public price - 4CoreWeave (W&B Inference)CoreWeave
Lowest measured price on gpt-oss-120b, but slow output in Artificial Analysis tests
Best for Batch jobs on gpt-oss-120b where cost matters more than speedYou give up Slower output than most hosts on this list
80ChecklistNo public price - 5Novita AINovita
Cheap, broad catalog of open models including Kimi K3 and GLM-5.3
Best for Trying many open models cheaply from one accountYou give up Same price as Crusoe but scores lower (80 against 90) and replies a little slower
80ChecklistPay as you go - 6BasetenBaseten
Fast first responses for production apps, with dedicated GPUs when you outgrow shared APIs
Best for Production apps where the first word of a reply must arrive fastYou give up Twice Crusoe's per-token price for the same open model
80ChecklistPay as you go - 7Nebius Token FactoryNebius
Fast, mid-priced open models; useful if you want an EU-based host
Best for European teams that want fast open model output from an EU-based hostYou give up Costs several times DeepInfra's rate; price taken from OpenRouter
80ChecklistNo public price - 8Amazon BedrockAmazon
For AWS teams wanting Claude, GPT and open models in one place; slower on open models
Best for Companies on AWS that want Claude Opus 5.5 and Kimi K3 in their own accountYou give up Slow on open models (87 tokens a second); US price not shown
80ChecklistPay as you go - 9Google Vertex AIGoogle
For Google Cloud teams wanting Gemini, Claude and open models with one bill
Best for Google Cloud teams that want Gemini 4 Argon and Claude Opus 5.5 togetherYou give up Was down in one OpenRouter uptime check; output slower than the median host
70ChecklistPay as you go - 10GroqGroq
Very fast replies for latency-sensitive apps, but only a few small open models
Best for Voice, chat or agent apps where fast replies matter more than model choiceYou give up Only gpt-oss and Llama models, which rank low on a public leaderboard
70ChecklistPay as you go - 11Together AITogether AI
Wide open-model catalog plus fine-tuning and GPU rental in one account
Best for Teams that plan to fine-tune open models as well as call themYou give up Slow in speed tests, and its GPUs cost more per hour than DeepInfra's
70ChecklistPay as you go - 12Fireworks AIFireworks AI
Strong pick for top open models like Kimi K3 and GLM-5.3 with cheap fine-tuning
Best for Teams fine-tuning smaller open models on a budgetYou give up On-demand GPUs at $8 an hour, twice Together's rate; token prices only in its docs
70ChecklistPay as you go - 13ParasailParasail
Smaller host with the main open models; output price is above most rivals
Best for A second source for Kimi K3 and the main open modelsYou give up Scores below hosts that charge less, like Crusoe and DeepInfra
60ChecklistPay as you go - 14SambaNova CloudSambaNova
Very fast output, but a small, older model list and high output price
Best for Apps that need very fast output on the older models it offersYou give up First reply is slower than most hosts here, despite the fast output
60ChecklistNo public price - 15Scaleway Generative APIsScaleway
For EU teams who want a European host and a small free allowance
Best for Trying open models from a European host at no cost at firstYou give up Its data policy is unconfirmed, and prices are at the high end of this list
50ChecklistFree - 16Cerebras InferenceCerebras
The fastest measured host, worth it only if raw speed matters more than model choice
Best for Fast agent loops on gpt-oss-120b or Qwen3.8 27BYou give up Only two models, and its listed prices are among the highest here
50ChecklistNo public price - 17OpenRouterOpenRouter
One key for almost every model and host, with automatic fallback between providers
Best for Comparing many models with one key, including free ones to startYou give up Adds a 5.5% fee on Standard; the free plan caps you at 50 requests a day
Not scored yetFree - 18Hugging Face Inference ProvidersHugging Face
Try many hosts through your Hugging Face account at the hosts' own prices
Best for Hugging Face users who want to try several hosts without new accountsYou give up Free accounts must buy credits; not scored yet: no independent test
Not scored yetFrom $2/mo - 19Cloudflare Workers AICloudflare
Handy if your app already runs on Cloudflare; small free daily allowance
Best for Small apps on Cloudflare that fit inside a free daily allowanceYou give up Not scored yet, and its paid price matches the highest on this list
Not scored yetFree
| # | Name | Score | What placed it | Price | Best for | You give up |
|---|---|---|---|---|---|---|
| 1 | DeepInfraDeepInfra · Closed | 90Our checklist | 8th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 232 t/s, 0.72s first chunk. checklist: job 2, model 1, value 2, privacy 2, limits 2 = 9; AA 232 t/s (Turbo 396). Report a mistake | API $0.037 in / $0.17 out per 1M tokens | Cheap, high-volume use of popular open models like DeepSeek, Qwen and GLM | No free tier; you need a card or prepay before your first call |
| 2 | Crusoe Managed InferenceCrusoe · Closed | 90Our checklist | 7th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 236 t/s, 0.42s first chunk. checklist: job 2, model 1, value 2, privacy 2, limits 2 = 9; AA 236 t/s. Report a mistake | API $0.05 in / $0.25 out per 1M tokens | Chat apps on open models where a fast first reply matters | Costs a bit more per token than DeepInfra at the same score |
| 3 | Azure AI FoundryMicrosoft · Closed | 90Our checklist | 4th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 361 t/s, 1.62s first chunk. checklist: job 2, model 2, value 1, privacy 2, limits 2 = 9; AA 361 t/s; offers Claude Fable 5.1 (preview); value from Azure retail price $0.15/$0.60 Report a mistake | Not confirmed | Companies already buying from Microsoft that need big-name models with fast output | Slowest first reply on this list, though output is fast once it starts |
| 4 | CoreWeave (W&B Inference)CoreWeave · Closed | 80Our checklist | Slowest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 54 t/s, 1.21s first chunk. checklist: job 1, model 1, value 2, privacy 2, limits 2 = 8; AA 54 t/s (slow); $0.03/$0.17 via OpenRouter listing; offers DeepSeek V4.1 Flash Report a mistake | Not confirmed | Batch jobs on gpt-oss-120b where cost matters more than speed | Slower output than most hosts on this list |
| 5 | Novita AINovita · Closed | 80Our checklist | 9th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 224 t/s, 0.90s first chunk. checklist: job 1, model 1, value 2, privacy 2, limits 2 = 8; AA 224 t/s, just under median 232. Report a mistake | API $0.05 in / $0.25 out per 1M tokens | Trying many open models cheaply from one account | Same price as Crusoe but scores lower (80 against 90) and replies a little slower |
| 6 | BasetenBaseten · Closed | 80Our checklist | 6th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 238 t/s, 0.19s first chunk. checklist: job 2, model 1, value 1, privacy 2, limits 2 = 8; AA 238 t/s, fastest first chunk 0.19s; offers Kimi K3 Report a mistake | API $0.10 in / $0.50 out per 1M tokens | Production apps where the first word of a reply must arrive fast | Twice Crusoe's per-token price for the same open model |
| 7 | Nebius Token FactoryNebius · Closed | 80Our checklist | 5th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 360 t/s, 0.96s first chunk. checklist: job 2, model 1, value 1, privacy 2, limits 2 = 8; AA 360 t/s (Base); $0.15/$0.60 per OpenRouter provider page; offers GLM-5.3 Report a mistake | Not confirmed | European teams that want fast open model output from an EU-based host | Costs several times DeepInfra's rate; price taken from OpenRouter |
| 8 | Amazon BedrockAmazon · Closed | 80Our checklist | 13th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 87 t/s, 0.94s first chunk. checklist: job 1, model 2, value 1, privacy 2, limits 2 = 8; AA 87 t/s; offers Claude Opus 5.5 and Kimi K3; value from $0.15/$0.60 via OpenRouter Report a mistake | API $0.15 in / $0.62 out per 1M tokens (Sydney; US price not shown) | Companies on AWS that want Claude Opus 5.5 and Kimi K3 in their own account | Slow on open models (87 tokens a second); US price not shown |
| 9 | Google Vertex AIGoogle · Closed | 70Our checklist | 11th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 136 t/s, 0.75s first chunk. checklist: job 1, model 2, value 1, privacy 2, limits 1 = 7; AA 136 t/s; offers Gemini 4 Argon and Claude Opus 5.5; limits 1: endpoint down in OpenRouter 30-min uptime check Report a mistake | API $0.09 in / $0.36 out per 1M tokens | Google Cloud teams that want Gemini 4 Argon and Claude Opus 5.5 together | Was down in one OpenRouter uptime check; output slower than the median host |
| 10 | GroqGroq · Closed | 70Our checklist | 3rd fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 474 t/s, 0.70s first chunk. checklist: job 2, model 0, value 1, privacy 2, limits 2 = 7; AA 474 t/s. Report a mistake | API $0.15 in / $0.60 out per 1M tokens | Voice, chat or agent apps where fast replies matter more than model choice | Only gpt-oss and Llama models, which rank low on a public leaderboard |
| 11 | Together AITogether AI · Closed | 70Our checklist | 14th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 74 t/s, 0.58s first chunk. checklist: job 1, model 1, value 1, privacy 2, limits 2 = 7; AA 74 t/s on gpt-oss-120b; offers Kimi K3 and GLM-5.3 Report a mistake | API $0.15 in / $0.60 out per 1M tokens | Teams that plan to fine-tune open models as well as call them | Slow in speed tests, and its GPUs cost more per hour than DeepInfra's |
| 12 | Fireworks AIFireworks AI · Closed | 70Our checklist | About 91 tokens (word pieces) a second on Kimi K3 (Artificial Analysis)What it rests onTheir figure: AA Kimi K3 (max) 91 t/s, 1.03s first chunk. checklist: job 1, model 1, value 1, privacy 2, limits 2 = 7; not measured on gpt-oss-120b; AA measures it on Kimi K3 at 91 t/s Report a mistake | API $0.15 in / $0.60 out per 1M tokens | Teams fine-tuning smaller open models on a budget | On-demand GPUs at $8 an hour, twice Together's rate; token prices only in its docs |
| 13 | ParasailParasail · Closed | 60Our checklist | 10th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 156 t/s, 0.82s first chunk. checklist: job 1, model 1, value 0, privacy 2, limits 2 = 6; AA 156 t/s; offers Kimi K3 Report a mistake | API $0.10 in / $0.75 out per 1M tokens | A second source for Kimi K3 and the main open models | Scores below hosts that charge less, like Crusoe and DeepInfra |
| 14 | SambaNova CloudSambaNova · Closed | 60Our checklist | 2nd fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 706 t/s, 1.17s first chunk. checklist: job 2, model 0, value 0, privacy 2, limits 2 = 6; AA 706 t/s; $0.14/$0.95 per OpenRouter provider page; older models only Report a mistake | Not confirmed | Apps that need very fast output on the older models it offers | First reply is slower than most hosts here, despite the fast output |
| 15 | Scaleway Generative APIsScaleway · Closed | 50Our checklist | 12th fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 106 t/s, 1.37s first chunk. checklist: job 1, model 1, value 0, privacy 1, limits 2 = 5; AA 106 t/s; privacy 1: data policy not confirmed today Report a mistake | API EUR 0.15 in / EUR 0.60 out per 1M tokens; first 1M tokens free | Trying open models from a European host at no cost at first | Its data policy is unconfirmed, and prices are at the high end of this list |
| 16 | Cerebras InferenceCerebras · Closed | 50Our checklist | Fastest of 15 hosts we list, same model (Artificial Analysis)What it rests onTheir figure: AA gpt-oss-120b 1,738 t/s, 0.42s first chunk. checklist: job 2, model 0, value 0, privacy 2, limits 1 = 5; AA 1,738 t/s, fastest measured; $0.35/$0.75 per OpenRouter; only gpt-oss-120b and Qwen3.8 27B Report a mistake | Not confirmed | Fast agent loops on gpt-oss-120b or Qwen3.8 27B | Only two models, and its listed prices are among the highest here |
| 17 | OpenRouterOpenRouter · Closed | Not scored | Can send gpt-oss-120b requests to 24 places that run itNotesTheir figure: gpt-oss-120b served by 24 provider endpoints. no independent speed test of the router itself in the last 90 days Report a mistake | Free plan 50 requests/day; Standard 5.5% fee; gpt-oss-120b from $0.03 in / $0.15 out per 1M | Comparing many models with one key, including free ones to start | Adds a 5.5% fee on Standard; the free plan caps you at 50 requests a day |
| 18 | Hugging Face Inference ProvidersHugging Face · Closed | Not scored | Can send requests to 18 hosting companies it listsNotesTheir figure: routes to 18 listed providers. no independent speed test of the router itself in the last 90 days Report a mistake | Provider rates with no markup; PRO gets $2/mo credits | Hugging Face users who want to try several hosts without new accounts | Free accounts must buy credits; not scored yet: no independent test |
| 19 | Cloudflare Workers AICloudflare · Closed | Not scored | Offers gpt-oss-120b, GLM-5.3 and DeepSeek V4NotesTheir figure: offers gpt-oss-120b, GLM-5.3, DeepSeek V4. not in Artificial Analysis provider benchmarks; no independent test in last 90 days Report a mistake | API $0.35 in / $0.75 out per 1M tokens; 10,000 free Neurons/day | Small apps on Cloudflare that fit inside a free daily allowance | Not scored yet, and its paid price matches the highest on this list |
Nothing matches all of that. , or try everything.
How we ranked this
Order. Highest score first; on a tie, the better price for the same model.
How we score this list. Our checklist: independent speed and reliability measurements, which models they offer, price, data policy and limits. The full rules are on How we score. Open "How it was scored" on any row to see the working.
What we did not do. We did not run any of these ourselves. Every number comes from the public source linked on its row, checked on 18 Oct 2026.
Spot a mistake? A wrong rank, price or link: tell us and we will check it against the source. Corrections are dated in News.
Reviews are drafted with help from AI models, including Anthropic's Claude. Model scores come only from outside leaderboards and independent evaluations, by a written rule.