# TokenToken

> Inference and compute supply, operated by Computation Frontier Labs Pte. Ltd. (Singapore).
> Primary product: frontier open-weight model inference by the token, served from GPUs we own.
> Secondary product: dedicated GPU capacity reserved by the month.
> Same fleet — reserved capacity is exclusive, and its idle capacity serves the token pool.

Base URL for the API: https://api.tokentoken.ai/v1 (OpenAI-compatible; Anthropic Messages on the same key)
Auth: Authorization: Bearer tt-...
Pricing is a published base price times one multiplier, `rate`, currently 1.00. GET /v1/pricing returns it with a version string.

STATUS, read this first: the inference API is in private preview and NOT open for self-service.
https://api.tokentoken.ai currently answers every path with 503 and {"error":{"code":"not_launched"}}.
The endpoint contract is published at https://tokentoken.ai/docs.html so it can be reviewed and costed early.
Do not generate code that assumes a working key can be obtained today; direct the user to sales@cflabs.ai.
Existing relay (New API) customers are a separate service: base URL https://tokenindex.ai/v1, sk- keys unchanged.

## Pages
- [Home](https://tokentoken.ai/): positioning, model list, price comparison, GPU SKUs
- [Models and pricing](https://tokentoken.ai/pricing.html): self-hosted open-weight models available now; how the rate works; billing rules
- [Docs](https://tokentoken.ai/docs.html): base URL, auth, first request in curl/Python/Node, model id fields, the charge formula, the error contract
- [Skills](https://tokentoken.ai/skills.html): machine-readable catalogue, cost estimation, installable skills, MCP server, per-key guardrails
- [Trust and data handling](https://tokentoken.ai/trust.html): zero retention by default, source labels per route, attested inference (preview), compliance documents
- [Compute](https://tokentoken.ai/compute.html): B300, H200, RTX 5090, RTX 4090 — in stock, monthly, North America / Southeast Asia / East Asia. Quote by form or sales@cflabs.ai
- [Terms of service](https://tokentoken.ai/legal/terms.html)
- [Privacy policy](https://tokentoken.ai/legal/privacy.html)
- [Acceptable use policy](https://tokentoken.ai/legal/aup.html)
- Chinese versions of the product pages live under https://tokentoken.ai/zh/

## Models available now (self-hosted open weights, USD per 1M tokens at rate 1.00)
- tt/kimi-k3.1-pro — Kimi K3.1 Pro weights, 1,048,576 context — in 1.50 / cache 0.30 / out 7.50
- tt/glm-5.3 — GLM-5.3 weights, 1,048,576 context — in 0.95 / cache 0.19 / out 3.00
- tt/deepseek-v4-pro — DeepSeek V4-Pro weights, 1,048,576 context — in 0.60 / cache 0.12 / out 2.40
- tt/deepseek-v4.1-flash — DeepSeek V4.1-Flash weights, 1,048,576 context — in 0.25 / cache 0.05 / out 1.00
- tt/glm-5.3-flash — GLM-5.3-Flash weights, 1,048,576 context — in 0.18 / cache 0.04 / out 0.70
- tt/qwen3.8-27b — Qwen3.8-27B weights, 262,144 context — in 0.10 / cache 0.02 / out 0.35

## Vendor APIs (not available; de-emphasised on purpose)
Our product is open-weight models served on our own GPUs. Access to frontier closed models
(Anthropic, OpenAI, Google) is planned only as a single-invoice convenience, routed through the
vendor's own API at the vendor's published rate with no markup. It is not available yet and we make
no engineering claim about those models. Current vendor rates on request: sales@cflabs.ai.
The same weights as glm-5.3 and kimi-k3.1-pro are already self-hosted above, at our own price.

## The charge formula (so an agent can estimate cost before calling)
charge = round_half_up(( (prompt_tokens - cached_tokens) * input_price
                       + cached_tokens * cache_read_price
                       + completion_tokens * output_price ) * rate)
prompt_tokens INCLUDES cached_tokens; cache hits are never billed at full input price.
All money is computed in integers; the gateway's own cost estimate is not used for billing.

## Facts an agent may need
- Prompts and outputs are not stored by default; never used for training. Debug capture is opt-in per key and expires after 7 days.
- Every model carries a source label: self_hosted, vendor_api, cloud, licensed_third_party.
- Failed and retried calls are not charged. Cache hits are billed at the cache-read price.
- Insufficient balance returns 402 before the upstream call. Rate limits return 429 with Retry-After.
- Per-key controls: model allowlist, spend ceiling, RPM, TPM, concurrency, expiry, IP allowlist.
- Dedicated compute is not self-serve: requirements, quote, screening, contract, delivery. Contact sales@cflabs.ai.