Skip to main content
SGL Grid is usage-based: you pay per request, using the model’s documented billing unit. There is no subscription. Chat and embedding workloads use token/input units. Release-gated transcription uses exact sample-derived audio duration and has a $0.0001 minimum.

How cost is calculated

  • Each model has a per-token price (input and output).
  • Cost for a request = tokens × the model's rate.
  • Credits are charged the exact amount after inference (on real token usage). x402 charges at request time based on the model and your max_tokens.
For the transcription candidate, cost is max(100, ceil(sample_count × 100 / 16000)) / 1000000 USD, using integer micro-USDC arithmetic and a $0.0001 minimum. A 4.2-second clip costs $0.000420. Duration stays exact; only the charge rounds up to micro-USDC. The quote comes from the authenticated sample count, not a declared or node-reported duration. See Speech-to-text transcription. The live per-model price is shown in the Models view and the models API:

Per-provider pricing

Most nodes serve at the platform suggested rate, but operators may set their own per-token price within a band (suggested × 0.5 to × 5). Image, speech and transcription models use a band of × 0.5 to × 2, see below. You can compare the nodes serving a model — each with its effective price, cheapest first — and optionally pin one:
Pass the chosen node_id as node on your chat request (or the node= argument in the SDKs) to use a specific provider; omit it and the grid routes for you.

Capping price (max_price)

Don’t want to pick manually? Pass max_price (blended USD per 1M tokens) on your chat/reserve request — the grid only routes to nodes at or under that rate, and the request is never billed above it. Works in every routing mode. It’s a typed option in the SDKs (max_price / maxPrice).
If no node qualifies under the cap, the request returns price_cap_unmet instead of overcharging you.

Image, speech and transcription prices

Providers can set their own price for image, text to speech and transcription models too. A provider picks a multiplier from 0.5× to 2× of the suggested rate, and the price you see is always the exact price you pay: it is quoted when you reserve, locked in for that request, and charged to the micro-dollar. A provider changing their price mid-request never changes what you owe. The provider keeps 80%.
  • The reserve response shows the quote. When a provider has set a custom price it also carries price_mult_ppm (parts per million, 1000000 is the suggested rate) and reference_price_usd.
  • Compare providers with GET /v1/providers?model=<id>: image, speech and transcription entries list each node’s price (price_per_mp_step, price_per_byte or price_per_audio_second), cheapest first.
  • Use max_price (USD for the whole request) on image, speech and transcription reserves. If the node you are given costs more, you get price_cap_unmet and nothing is charged.

Where the money goes

Each paid inference is split across the people who make the network work: So usage directly rewards operators and stakers, and continuously buys + burns $SGL.

Keeping costs down

  • Pick the smallest model that does the job (see Models).
  • Set a sensible max_tokens.
  • Use credits for steady usage (exact billing) and x402 for occasional/agent use.

Tracking spend

The dashboard’s API usage view shows your requests, tokens, spend, and a per-model breakdown over time. See Billing.