Illustrative. Generated with our own models on our own GPUs.
Kimi K3.1 Pro, GLM-5.3, DeepSeek V4 and Qwen 3.8 on an OpenAI-compatible endpoint. Run on hardware we own and operate, with the price published in full.
1M context · 99.9% monthly availability target · zero retention by default · rate 1.00
# OpenAI-compatible. Change the base URL. from openai import OpenAI client = OpenAI( base_url="https://api.tokentoken.ai/v1", api_key="tt-...", ) r = client.chat.completions.create( model="tt/kimi-k3.1-pro", messages=[{"role": "user", "content": "hi"}], )
Anthropic Messages supported on the same key.
Illustrative. Generated with our own models on our own GPUs.
One team owns the kernels, the scheduler, the fleet and the pager. That is what makes a route stay up under load — and what lets us price it competitively instead of marking up someone else's quota.
One key, one endpoint. The gateway authorizes, reserves budget and routes; the fleet produces the tokens.
Self-hosted routes are live. Vendor-API routes are being onboarded and are marked coming soon in the catalogue and in GET /v1/models.
Every failure path is detect → remediate → verify, and each one is held by an automated gate that injects the fault and asserts the remediation fires. Detection without remediation counts as unfixed.
One directory per release, so the previous one is the rollback. Read-only code split from runtime state. One build SHA to every node. On call, with a named owner per service.
Admission control, backpressure, retry budgets, slow-node demotion and overload shedding — so your p99 does not collapse when someone else's traffic spikes.
Attention and GEMM kernels tuned per card, quantization, speculative decoding, continuous batching. This is where throughput per GPU comes from, and we do the work ourselves.
Not resold quota. Reserved hardware is yours alone, and its idle time feeds the token pool — so supply and price move with our own engineering, not with a vendor's price list.
Every route declares its source in the API response. Prompts and outputs are not stored. Selected routes return a signed attestation your client checks before the prompt leaves.
Every price is a base price times one public multiplier, rate, currently 1.00. It is on this site, in the console and in the API — the same number we bill.
A named contact for business accounts, a request id on every response that ties a call to its line on the invoice, and status and incident notices you do not have to ask for.
Frontier open weights on our own hardware. Vendor APIs are being onboarded.
| Model | Weights | Context | Input /M | Output /M | Status |
|---|---|---|---|---|---|
| tt/kimi-k3.1-pro | Kimi K3.1 Pro | 1M | $1.50 | $7.50 | Available |
| tt/glm-5.3 | GLM-5.3 | 1M | $0.95 | $3.00 | Available |
| tt/deepseek-v4-pro | DeepSeek V4-Pro | 1M | $0.60 | $2.40 | Available |
| tt/deepseek-v4.1-flash | DeepSeek V4.1-Flash | 1M | $0.25 | $1.00 | Available |
| tt/glm-5.3-flash | GLM-5.3-Flash | 1M | $0.18 | $0.70 | Available |
| tt/qwen3.8-27b | Qwen3.8-27B | 262K | $0.10 | $0.35 | Available |
| claude-opus-5-5 | Anthropic API | 1M | — | — | Coming soon |
| gpt-6-astra | OpenAI API | — | — | — | Coming soon |
| gemini-4-argon | Google API | — | — | — | Coming soon |
Prices at rate 1.00. Vendor-API models will list at the vendor's published rate while we onboard them.
Where our self-hosted routes sit against each vendor's own published rate.
| Model | Source | Input /M | Output /M |
|---|---|---|---|
| tt/deepseek-v4-pro | TokenToken self-hosted | $0.60 | $2.40 |
| tt/glm-5.3 | TokenToken self-hosted | $0.95 | $3.00 |
| tt/kimi-k3.1-pro | TokenToken self-hosted | $1.50 | $7.50 |
| glm-5.3 | Z.ai published rate | $1.40 | $4.40 |
| gemini-4-argon | Google published rate | $2.00 | $10.00 |
| claude-opus-5-5 | Anthropic published rate | $4.00 | $20.00 |
| gpt-6-astra | OpenAI published rate | $10.00 | $50.00 |
Vendor rates are each vendor's own published standard-tier price for requests inside their short-context band, read 6 November 2026; promotional rates and long-context bands differ. Our prices are shown at rate 1.00.
Most teams start on the API and pay per token. When your load is steady enough to keep a card busy, reserve one: B300 in stock, by the month, power and bandwidth included. We quote within one business day.
What fits without splitting the model across cards.
HBM3e on B300 and H200; GDDR7 and GDDR6X on the RTX cards. Bars are to scale.
Discovery, cost and limits are endpoints — nothing here needs a human to read docs.
A model allowlist, a spend cap and a rate limit sit on the key itself, so an agent that loops cannot outspend its budget.
# Catalogue, prices and limits as data curl https://api.tokentoken.ai/v1/models curl https://api.tokentoken.ai/v1/pricing # What will this call cost? curl -X POST .../v1/cost/estimate \ -d '{"model":"tt/glm-5.3","input_tokens":12000}' # Agent-readable site index curl https://tokentoken.ai/llms.txt
SSE, parallel tool calls and JSON-constrained output on every self-hosted route.
OpenAI and Anthropic client libraries work unchanged — base URL and key.
Copy-paste skill files for coding agents, with the limits already set.
A pinned tt/ id keeps serving the same weights and quantization.
Prompts and outputs are not stored. Billing metadata is. Debug capture is opt-in per key and expires on its own.
No setting changes this and no discount is offered for allowing it.
Self-hosted, vendor API, cloud or licensed third party — in the API response, not only on this page. No account pools.
Selected routes run in a hardware TEE and return an attestation your client verifies before the prompt leaves, plus a signed usage receipt.
Inference can be contracted to stay in North America, Southeast Asia or East Asia.
Authorization headers are dropped at the entry point, not redacted afterwards.
DPA with SCCs and UK Addendum, sub-processor list with 30 days' notice, security overview.