Fireworks AI is a fast inference platform for open-weight models — Llama, DeepSeek, Qwen, Mixtral and others — served through an OpenAI-compatible API. You pay per token (scaled by model size) for serverless, or per GPU-hour for dedicated capacity. It's a go-to when you want open models cheaper than closed frontier ones, with low latency.
| Model class | Price / 1M tokens | Examples |
|---|---|---|
| Small (≤16B) | ~$0.20 | Llama 8B, small Qwen |
| Mid (16–80B) | ~$0.90 | Llama 70B, Qwen 72B |
| Large MoE / 100B+ | ~$1.20–$3.00 | DeepSeek V3, Llama 405B |
| Dedicated GPU (on-demand) | per GPU-hour | reserved throughput, e.g. H100/H200 |
| Fine-tuning & hosting | per token / per GPU | train + serve your own variant |
Fireworks gives you two billing models. Serverless charges per token with no idle cost — ideal for spiky or low-volume traffic, since you pay nothing between requests. Dedicated on-demand GPUs charge by the hour but give you reserved throughput and the lowest latency, which becomes cheaper than per-token once your volume is high and steady enough to keep the GPU busy. The crossover is a utilisation question: estimate your token volume, then compare per-token serverless against the hourly GPU rate with the self-hosted vs API calculator and the GPU cloud cost calculator.
New Fireworks accounts usually get a small free credit to test, along with workable rate limits on serverless models, rather than a permanent free tier. After the credit you draw down a billed balance per token. For experimentation it's plenty; for production, price your real token mix with the calculators below.
1. Sign up at fireworks.ai and open the dashboard.
2. Claim any free credit and add billing for production use.
3. Go to API Keys and create a key.
4. Use the OpenAI-compatible endpoint — just change the base URL and model name.
Together AI (see the Together guide) is the closest competitor — same open-model, per-token model with similar pricing; benchmark both for your specific model. Replicate (see Replicate) is stronger for image/video and cold-start one-offs. Groq wins on raw speed for supported models. Against closed models, running an open model on Fireworks is typically several times cheaper per token than OpenAI or Claude — the trade is choosing and owning the open model. Rank everything for your token mix with the cheapest LLM API tool.
Per token by model size: ~$0.20/1M for small models, ~$0.90/1M for mid, ~$1.20–$3/1M for large MoE models (reference, July 2026). Dedicated GPUs are billed per hour.
Yes — use the OpenAI SDK, point the base URL at api.fireworks.ai and set the model to the Fireworks model path.
When your traffic is high and steady enough to keep a GPU busy — reserved throughput beats per-token cost above the utilisation crossover. Low or spiky traffic stays cheaper on serverless.
Yes — Fireworks supports fine-tuning open models and serving the result, billed per training token and per inference token. Compare with the fine-tuning calculator before committing.
Not affiliated with Fireworks AI. Prices are reference estimates — always verify on the official pricing page.