What for: one of the cheapest OpenAI-compatible inference clouds for open models — pay per token, with no minimum spend.

What it is and who it's for

DeepInfra is an inference cloud for open models built around the lowest possible price per token and dead-simple usage: one OpenAI-compatible API, pay as you go, no minimum spend and no complicated tiers. It's for developers and startups who need to run Llama/Qwen/DeepSeek cheaply in production without babysitting GPUs. What sets it apart is that they run their own hardware in their own data centers, which is how they get to some of the lowest rates on the market.

Key capabilities

50+ open models (Llama, Qwen, DeepSeek, Mistral and others) at a low price per token.
An OpenAI-compatible API — migrate by swapping the base URL and key.
Not just LLMs: embeddings, STT (Whisper) and image models.
Dedicated deployments for stable latency in high-load scenarios.
One of the providers behind Hugging Face Inference Providers.
A starter free credit with no card required.

Get started in 5 minutes

Sign up at deepinfra.com and grab an API token.
Call a model through the OpenAI-compatible endpoint:

curl https://api.deepinfra.com/v1/openai/chat/completions \
  -H "Authorization: Bearer $DEEPINFRA_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model":"meta-llama/Llama-3.3-70B-Instruct","messages":[{"role":"user","content":"Hello"}]}'

Compare the per-token price against other providers for your model — it often comes out cheaper.

When to use it and when not to

✅ Use it if the lowest price per token on open models and a simple OpenAI-compatible call are what matter most.
✅ Use it if you want cheap STT/embeddings alongside your LLM from one place.
❌ Skip it → go with "Together AI"/"Fireworks" if you need rich fine-tuning and production deployments with guarantees.
❌ Skip it → go with "OpenRouter" if you care about top proprietary models and one bill across hundreds of providers.

The honest price

Billed by tokens (some models by inference time); the rates are among the lowest, but exact figures and model availability change. There's a starter credit. Check the provider for current numbers.

Pitfalls

Cheap doesn't always mean fast: for tight latency requirements, take a dedicated deployment.
The catalogue is open models; there's no proprietary frontier tier.
Whether a given model and version is available changes over time — pin it in your config.
It's the cloud: sensitive data leaves your perimeter — check your compliance.

🤖 Prompt accelerator

"I want to cut my inference costs and I'm considering DeepInfra. For , pick an open model, give me a sample OpenAI-compatible request in <Python/Node>, work out the cost per 1M tokens and compare it to my current provider. Tell me when it's worth moving to a dedicated deployment for latency and how to measure time-to-first-token."