What is Fireworks AI
Fireworks AI provides serverless and dedicated inference, fine-tuning and managed training for open and proprietary AI models.
Serverless inference
As of September 7, 2026, new accounts receive US$1 in free credits. Standard serverless base-model pricing per million tokens is US$0.10 for models below 4B parameters, US$0.20 for 4B to 16B, US$0.90 above 16B, US$0.50 for mixture-of-experts models up to 56B active parameters and US$1.20 for 56.1B to 176B. Batch inference is 50% below the standard rate. Priority and Fast tiers provide different performance and pricing.
Published embedding rates start at US$0.008 per million tokens for models up to 150M parameters and US$0.016 for 150M to 350M. Current headline rates include GLM 5.3 at US$1.40 input, US$0.26 cached input and US$4.40 output; Kimi K3 at US$3, US$0.30 and US$15; and DeepSeek V4 Flash (0731) at US$0.22, US$0.007 and US$0.66 per million tokens.
Dedicated compute and training
On-demand H100 and H200 deployments cost US$8 per GPU hour, B200 costs US$13, B300 costs US$15 and GB300 costs US$20, billed per second. US-only serverless models launched from September 1, 2026 generally carry a 50% premium, with documented exceptions. Managed training and reserved capacity are priced separately.
Official sources: Fireworks pricing and serverless pricing documentation.