Overview
Serverless inference is priced per token. For how Standard, Priority, Fast, and Ultra-fast (Preview) modes work and how to select one, see Serverless Modes. Every text or vision request is billed across three dimensions:- Input tokens — what you send to the model.
- Cached input tokens — input tokens served from prompt cache, priced lower.
- Output tokens — what the model generates.
How pricing works
- Prices below are per 1 million tokens in US dollars.
- Batch inference is billed at 50% of serverless pricing on both input and output. See Batch inference.
Text and vision models
Per-model pricing for headline models. Fast and Ultra-fast (Preview) variants appear as adjacent rows. In each Standard or Priority cell, prices are input / cached input / output (USD per 1M tokens), in that order.
— in the Priority column means Priority is not available for that model. This pricing table is the source of truth for Priority availability.
Muse Spark Open runs on US-only capacity. The 50% US-only premium is waived, so the rates in that row are what you pay.
Decision models
Rune 26B-A4B is a System One decision model. It reads the input once and returns calibrated probabilities for a set of typed questions, and it never generates output tokens. You pay for input only, so the rates below cover the full cost of a request.
Cached input is 20% of the input rate. Call it with
model: "accounts/fireworks/models/rune"; see the System One decisions API for the request shape.
Rune runs on US capacity. There is no US-capacity premium, so the rates above are what you pay.
Other base models — by size and architecture
For any text or vision model not listed individually, pricing is set by parameter count and architecture. These size-based prices apply uniformly to input and output (no separate cached-input rate):Embeddings
Embeddings are billed per 1M input tokens.Notes
- For account-level controls (spend tiers, monthly spend limits, on-demand GPU quotas), see Account quotas.