Fireworks AI

Fireworks AI is a hosted inference platform for open models. The Fireworks API is OpenAI compatible: create an API key in the Fireworks dashboard, export it as FIREWORKS_API_KEY, point the OpenAI Python or JavaScript client at https://api.fireworks.ai/inference/v1, and use model ids of the form accounts/fireworks/models/<name>. The same key also works with the Anthropic SDK through base URL https://api.fireworks.ai/inference, and Fireworks has its own Python SDK (pip install --pre fireworks ai), which its quickstart marks as alpha. Unlike OpenAI, the API also returns token usage in the last chunk of a streamed response.

Pricing, per the official pricing page on October 28, 2026: serverless inference is billed per input, output and cached token from prepaid credits, with optional auto reload and monthly spend limits. On demand deployments are billed per GPU second with no charge for start up time, for example $8.00 an hour for an H100 or H200, $13.00 for a B200, $15.00 for a B300 and $20.00 for a GB300; region restricted deployments cost 1.5 times as much. Managed LoRA and full fine tuning is billed per million training tokens (LoRA SFT starts at $0.50 per million for models up to 16B parameters), and fine tuned models are served at the base model's price.

For agent builders it plugs into the Vercel AI SDK through @ai-sdk/fireworks, into LiteLLM as a provider, and it is one of the providers OpenRouter routes to. Alternatives include Together AI, Groq, Baseten and Replicate.