# Limits & pricing

> **Beta.** The limits reflect the current beta configuration.

## Pricing

- **$0.09 per million input tokens** (`spinf-12b`, Gemma 4 12B).
- Processed tokens = each content once + for every question, the filled template and each answer's tokens after its
  first + the same for the empty floor, once per distinct question in the call (set `empty_floor: false` to skip it).
  No system prompt or other hidden tokens. For example, the [quickstart](/docs/quickstart) call processes 148 tokens
  (50 of content, 49 for its 4 questions, 49 for their floors) and is billed the 1,000-token minimum.
- Minimum **1,000 tokens per call**, counted on the whole call: send many questions, and many contents with
  `inputs`, in one call. Each model's minimum is `min_billed_tokens_per_call` in `GET /v1/models`.
- [Prompt-packs](/docs/prompt-packs) have no surcharge: their questions are billed like yours (content moderation:
  1,626 question tokens per content, plus the content once, shared with your own questions).
- The output (the scores) is free.
- A request needs **$0.10 of credit** to start. It is charged after scoring, on the tokens actually processed; a
  refused or failed request is not charged.
- New accounts start with **$4.50 of free credit** (50M tokens).

## Limits of the open service

| Limit | Value |
|---|---|
| Content per input | 4,096 tokens (8,192 under evaluation) |
| Questions per input | up to 10,000 (templates × combinations) |
| Tokens per question | 256 (the filled template) |
| Answers per question | up to 1,000, each 1–5 tokens |
| Placeholders per template | up to 5; up to 1,000 combinations per template |
| Processed tokens per request | up to 2M; request body up to 2 MB |
| Requests per minute | 600 |
| Concurrent requests | 1 per $100 of credit balance (at least 1); more on request |

The system has been tested with content up to 64K tokens and longer questions. In that configuration the current models become less effective, and it is still under investigation.

## Throughput

The on-demand service is built for throughput, not real time. It runs on elastic capacity, so the first request after an idle period may take a few seconds. After that, expect around 10,000 tokens per second.

## Real time: reserved endpoints

For decisions inside live flows (routing tickets as they arrive, moderating posts before they are published, agents),
we reserve an endpoint for you: capacity that is always warm and sized to your traffic. Calls are pinned to it with the
`endpoint` field of the request ([Scoring API](/docs/scoring-api)). A good way to start is to calibrate on your history
with the on-demand service, then move the same questions and thresholds to the reserved endpoint. Reserved endpoints
are priced per project: contact us at info@spinf.com.

## Large projects

For projects over 1T tokens we offer dedicated deployments and file-based (S3) batch analysis. The engine runs on datacenter-grade and consumer hardware (24 GB GPU minimum), and can be deployed on your premises when you need an extra level of privacy and security. Contact us at info@spinf.com.