Self-hosted models are priced by the token from prepaid credits. Dedicated GPUs are priced by the month and contracted. Every number on this page is a base price times the published rate multiplier.
Open weights we run ourselves. The tt/ id pins the weights, the version and the quantization.
| Model | Weights | Context | Input /M | Cache hit /M | Output /M | Served on |
|---|---|---|---|---|---|---|
| tt/kimi-k3.1-pro | Kimi K3.1 Pro | 1,048,576 | $1.50 | $0.30 | $7.50 | B300 |
| tt/glm-5.3 | GLM-5.3 | 1,048,576 | $0.95 | $0.19 | $3.00 | B300 / H200 |
| tt/deepseek-v4-pro | DeepSeek V4-Pro | 1,048,576 | $0.60 | $0.12 | $2.40 | B300 / H200 |
| tt/deepseek-v4.1-flash | DeepSeek V4.1-Flash | 1,048,576 | $0.25 | $0.05 | $1.00 | H200 |
| tt/glm-5.3-flash | GLM-5.3-Flash | 1,048,576 | $0.18 | $0.04 | $0.70 | H200 / 5090 |
| tt/qwen3.8-27b | Qwen3.8-27B | 262,144 | $0.10 | $0.02 | $0.35 | 5090 / 4090 |
USD per million tokens, at rate 1.00. Each model profile publishes its weight source, licence, revision and quantization. Other open-weight models deployed on request.
Our self-hosted rate against each vendor's own published rate.
Vendor rates are each vendor's published standard-tier price inside their short-context band, read 6 November 2026. Promotional rates and long-context bands differ — the tables on this page carry the detail.
Every published price is a base price times one number, the rate. It is 1.00 today and it is shown on this page, in the console and in the API.
We set the base prices from what it actually costs us to serve a token on our own fleet. When our optimization work lands — better kernels, higher batch efficiency, better cache hit rates — the rate is how that reaches your invoice.
Enterprise agreements can carry their own rate. Any change applies to calls made after it takes effect; your invoice records the rate version used for every request.
# The rate is data, not a footnote $ curl https://api.tokentoken.ai/v1/pricing { "rate": 1.00, "rate_version": "2026-11-06", "currency": "USD", "models": [ { "id": "tt/glm-5.3", "source": "self_hosted", "input": 0.95, "cache_read": 0.19, "output": 3.00 } ] }
| GPU | Memory | Regions | From /GPU·hr | Status |
|---|---|---|---|---|
| B300 | 288 GB | NA · SEA · EA | $6.50 | In stock |
| H200 | 141 GB | NA · SEA · EA | $2.60 | In stock |
| RTX 5090 | 32 GB | SEA · EA | $0.75 | In stock |
| RTX 4090 | 24 GB | SEA · EA | $0.45 | In stock |
Indicative, monthly commitment, power and bandwidth included. Longer terms carry a discount. Final price on quote.
Our product is open-weight models served on our own fleet. Access to frontier closed models — Anthropic, OpenAI, Google — is planned as a convenience for teams that want one invoice, routed through the vendor's own API under our commercial account and listed at the vendor's published rate. We do not mark those up, and we do not claim any engineering credit for them.
Not available yet. The open-weight routes above are what we run, optimize and stand behind — including the same weights as glm-5.3 and kimi-k3.1-pro, self-hosted, at our own price. If closed-model access matters to your evaluation, write to sales@cflabs.ai and we will send the current vendor rates in writing.
Top up by card, Alipay or WeChat Pay; bank transfer above $1,000. Credits do not expire. Auto top-up keeps production traffic from stopping on an empty balance.
An estimate is held when the request arrives and reconciled against the real token counts when it completes. An empty balance returns 402 before the call is made, never mid-stream.
A cached prefix is billed at the cache-read rate, not the full input rate. Reasoning and multimodal tokens are normalized per model so nothing is counted twice.
You are charged for the attempt that succeeded. Routing retries and failover are on us.
Per-key spend ceilings, model allowlists, RPM, TPM and concurrency limits. Set them in the console or through the API.
Monthly statements and per-top-up invoices from Computation Frontier Labs Pte. Ltd. Bank transfer and net terms for contracted accounts.