Atlas Inference

Pricing

Provisional per-model prices, and what free means during the beta.

Beta usage is free. No payment method is collected, no invoice is issued, and no balance accrues. The prices below are provisional: they are what your usage would cost, published so you can model it, and they may change before billing is enabled.

Provisional prices

Per million tokens, in USD.

ModelInputOutput
atlas-small-1$0.15$0.45
atlas-mid-1$0.50$1.50
atlas-large-1$2.00$6.00
atlas-code-1$0.40$1.20

GET /v1/models is authoritative — each entry carries a pricing object with provisional: true. Read it from your code rather than hard-coding this table.

What counts as a token

Inputbackend_reported_input_tokens

Everything you send: system, developer, user, and assistant messages, replayed tool calls and their results, and tool definitions.

Outputbackend_reported_output_tokens

Everything the model generates. Where a model emits a reasoning prelude, those tokens are counted inside this figure and priced as output — so a short answer can carry output tokens that were never delivered to you. Never add reasoning_tokens on top.

You are billed on what was generated, not on what was delivered. If your client disconnects mid-stream, the compute already happened, and the generated figure is what is metered. The delivered figure is recorded separately so you can see the difference — see Usage and cost.

Rate-limited and rejected requests write no usage row and cost nothing.

Prices are stamped on each request

Every usage row carries the price that was in force when the request ran. Cost figures are projected from that stamped price, never recomputed from today's published price.

So a price change does not retroactively reprice earlier usage, and an export you took last month still adds up to the same total today. This is also why the export carries input_price_per_mtok and output_price_per_mtok on every row: so you can recompute the cost columns rather than trust them.

Not priced during the beta

There is no charge for anything, but for completeness, these are not separately metered dimensions in any case: requests, streaming, tool calls, structured output, usage reads, or exports. Price is a function of input and output tokens only.

What is not offered yet

Batch submission and any discounted batch tier, prepaid credits, invoicing, committed-use discounts, and published enterprise pricing are all out of scope for the beta. When billing is enabled, the cumulative bound arrives with it as a prepaid balance; until then, rate limits are throughput protection and nothing caps cumulative usage.

On this page