New accounts: 20% off the first 20 deploys
Deploy your own dedicated AI server in minutes. Unlimited tokens on a GPU that only you use, far cheaper than per-token APIs once your workload gets busy.
$ curl $OSARAI_URL/v1/chat/completions \ -H "Authorization: Bearer $OSARAI_API_KEY" \ -d '{"model":"qwen3.8-27b","stream":true, "messages":[{"role":"user","content":"Hi"}]}'data: {"choices":[{"delta":{"content":"Hello!"}}]}data: {"choices":[{"delta":{"content":" How can I help?"}}]}data: [DONE]$ Compare
What API providers charge per million tokens for the same open model, and what your own Osarai GPU costs by the hour instead.
| Provider | Input | Output | Cached input | Precision | Context |
|---|---|---|---|---|---|
| PrimeIntellect Lowest | $0.020 | $1.08 | $0.018 | FP8 | 128K |
| Chutes Lowest | $0.039 | $1.25 | $0.004 | FP8 | 256K |
| Near AI Lowest | $0.040 | $1.35 | $0.018 | FP8 | 256K |
| StreamLake Lowest | $0.17 | $1.05 | $0.035 | – | 1M |
| Darkbloom Lowest | $0.035 | $1.50 | $0.018 | FP4 | 256K |
| Ionstream Lowest | $0.040 | $1.95 | $0.020 | FP8 | 256K |
| DeepInfra Lowest | $0.15 | $1.88 | $0.037 | BF16 | 256K |
| Modal Lowest | $0.040 | $2.30 | $0.020 | – | 256K |
| Tenstorrent Lowest | $0.15 | $2.00 | $0.050 | FP8 | 256K |
| Wafer Lowest | $0.060 | $2.30 | $0.059 | FP8 | 256K |
| DekaLLM Lowest | $0.040 | $2.50 | $0.020 | – | 256K |
| AkashML Lowest | $0.23 | $1.98 | $0.050 | FP8 | 256K |
| Parasail Lowest | $0.24 | $2.20 | $0.050 | FP8 | 256K |
| Reka Lowest | $0.15 | $2.50 | $0.050 | FP8 | 256K |
| Mancer 2 Lowest | $0.17 | $2.50 | – | FP8 | 256K |
| ModelRun Lowest | $0.20 | $2.50 | $0.090 | FP4 | 256K |
| Phala Lowest | $0.20 | $2.50 | $0.050 | – | 1M |
| Alibaba Lowest | $0.42 | $2.55 | $0.085 | – | 1M |
| CoreWeave Lowest | $0.40 | $3.00 | $0.15 | FP8 | 256K |
| Novita Lowest | $0.42 | $3.00 | $0.085 | – | 1M |
| Nebius Lowest | $0.45 | $3.00 | $0.10 | FP4 | 1M |
| Cerebras Lowest | $0.99 | $1.49 | $0.99 | FP16 | 64K |
| Cloudflare Lowest | $0.45 | $3.20 | $0.050 | – | 256K |
| Venice Lowest | $0.45 | $3.20 | – | FP8 | 256K |
Prices from OpenRouter's public provider list for Qwen3.8-27B, checked 6 November 2026: standard on-demand rates in US dollars, without batch discounts, promotions or taxes. Source
Savings
An Osarai instance costs the same per hour whether it writes a thousand tokens or a hundred million. Set your volume and see the monthly bill side by side.
Loading live price…
Estimate for 30 days of one always-on GPU at the current list price. Real throughput depends on prompt sizes and concurrency; heavy workloads may need more than one instance.
Models & pricing
Pick a model and a GPU. You get the whole card, with the model's full context window.
Qwen
Strong reasoning and coding in a model one GPU can serve fast.
OpenAI
OpenAI's open-weight model for fast, low-cost agents and chat.
OpenAI
OpenAI's largest open-weight model, on a 96 GB RTX PRO 6000.
Prices are per hour of a running GPU and follow the live GPU market. You see the exact price before every deploy.
Why dedicated
Your model runs on hardware reserved for you, behind an endpoint your code already knows how to call.
A whole GPU runs only your model. No noisy neighbours, no queue behind other customers.
No per-token bill and no token rate limits. Long contexts and agents cost the same hour.
Requests go to your own server. Osarai keeps token counts for your charts, not your text.
Point any OpenAI SDK, LangChain or your agent framework at your URL. Streaming included.
Top up on Base, Arbitrum, Optimism, BNB Chain, Polygon or Avalanche. No minimum deposit.
Idle instances shut down on their own, and you are warned before your balance runs out.
How it works
Sign in with Google, X, a passkey or your wallet, then send any amount of USDC to your deposit address.
Choose a model and a GPU. We rent the machine, load the model and show every step until it is ready.
Create a key and send requests to your endpoint, or chat with your model in the playground.
FAQ
You pay for the hours your GPU runs, however many tokens it produces. Busy workloads such as agents, coding assistants, RAG and batch jobs send millions of tokens a day; at that volume an hour of GPU costs far less than the same tokens priced per million. The calculator above shows where the break-even point sits for your volume.
Yes. Each instance is a whole GPU running only your model. No other customer's requests share it, so your latency does not depend on anyone else's traffic.
Qwen3.8-27B on an RTX 5090 or an RTX PRO 6000 (with image input on the PRO 6000), gpt-oss-20b on either GPU, and gpt-oss-120b on the RTX PRO 6000.
The endpoint speaks the OpenAI API, streaming included. Change the base URL and the API key in your OpenAI SDK, LangChain or agent framework and keep the rest of your code.
Your balance is prepaid in USDC on Base, Arbitrum, Optimism, BNB Smart Chain, Polygon or Avalanche. Each account gets its own deposit address and there is no minimum amount. A deposit shows as pending at once and is credited after the network confirms it.
As soon as you stop the instance. Auto-stop shuts it down after a period without requests, and the dashboard warns you before your balance runs out.
No. Requests are passed to your own GPU and the replies streamed back. Osarai records token counts for your usage charts, not the content of your prompts or replies.
Usually a few minutes: renting the machine, downloading the model and starting the server. The dashboard shows each step as it happens.
New accounts get 20% off the hourly price of their first 20 deploys within 30 days of signing up. Invited by a friend? Their link gives you 30% off your first 30 deploys instead. The discount lasts for each of those instances' whole life.
Yes. Share your invite link from the dashboard: friends who sign up through it get 30% off their first 30 deploys, and 10% of what they spend on GPUs comes back to your balance as credit.
今日から、自分専用のAIを。
Top up any amount, deploy in minutes and stop paying for every token.