DeepInfra

DeepInfra is an LLM inference API and GPU cloud for developers, serving open and proprietary models through one OpenAI-compatible endpoint, billed by usage.

DeepInfra hosts AI models behind an API, so a development team can run them in production without buying or operating GPUs. A request goes to a model DeepInfra serves, and the answer comes back in the format the OpenAI SDK already expects. Deep Infra Inc. builds and runs the service. The company was founded in September 2022, and its team has long experience building backend systems at very large scale. The serverless catalog lists each model with its context window and separate input and output token rates, plus a cached input rate on some. Text models come from DeepSeek, Moonshot AI's Kimi, Qwen, Meta's Llama, Google's Gemini and Gemma, NVIDIA Nemotron, Anthropic's Claude, Phi and Mistral. Beside text, it serves Voxtral speech transcription, FLUX image generation and bge embeddings through the OpenAI embeddings API. Developers get tool calling, JSON schema structured outputs, reasoning settings, image input and prompt caching through the chat API. The Batch API prices offline jobs lower, Service Tiers set priority per request, and dedicated GPU deployments get a private OpenAI-compatible endpoint. DeepInfra suits developers and product teams adding chat, agents, document reading or semantic search to an application without running a serving stack. Teams already coding against the OpenAI SDK can try open models without rewriting their client. Companies with steady traffic or a model of their own can move from shared serverless calls to rented GPUs. Billing is pay as you go, per token for serverless models and per GPU-hour for rentals, with no long-term contract. Larger GPU clusters are arranged directly with the sales team. A common job is an agent that calls a team's own functions and returns JSON held to a supplied schema. Another is a nightly run of document or embedding work sent through the Batch API, where cost matters more than speed. Vercel users can connect their DeepInfra account and export its API keys straight into Vercel. DeepInfra commits not to retain or log customer data beyond the time needed to serve a request. That term fits teams passing customer records through a model, with bulk jobs and Google and Anthropic models handled under separate rules.