Model access over API billed per token, not per seat — open and flagship models
Open and flagship language models over API, billed per token. An endpoint compatible with the OpenAI standard. Limits and usage accounting per key and per project.
When API access solves the task
API fits where computation does not have to sit inside the customer perimeter.
Fast start with no hardware
The scenario is validated in a day: a key, an endpoint, a first call. No accelerators to buy and no node to deploy.
Load is uneven
Requests rise and fall. A dedicated node sits idle; the API scales to demand.
Different models for different tasks
One model for classification, another for generation, a third for long context. Switching between models needs no code rewrite.
No personal data in the requests
If no personal data is transferred, an external call creates no localisation obligation. A perimeter is needed when data is present.
Which models are available
Two groups of models; selection is matched to the task and quality requirements.
How billing works
Billing is per token, not per seat. Limits and usage are visible per key and per project.
Rates and limits are fixed in the contract before work starts. Flagship-model terms are discussed separately.
When the perimeter is the better choice
API is not universal. In several cases on-premise deployment is required, and we say so before the contract.
- Personal data is in the requests. Then computation must run in the customer perimeter — a localisation requirement, not a price question.
- The largest closed model is required, not an open one. If quality depends on that closed model, a local replacement of equal quality is not guaranteed.
- Internal rules forbid external services. If the infosec policy excludes external calls, the API will not pass — a perimeter is needed.
What customers ask about API access
Which models can be connected?
Open models with open weights (Qwen, DeepSeek, GLM class) and flagship models from western vendors on request. Selection follows the task and acceptance criteria, not the price list.
Is there a per-user fee?
No. Billing is per token, with no per-seat subscription. The team can grow while the bill tracks real request volume.
Is the endpoint compatible with our code?
Yes. The interface is OpenAI-standard-compatible: the application is not tied to one model, and the model name changes without code rewrite.
How is spend limited?
Requests-per-minute and daily-volume caps are set to your load profile. Usage is visible per key and per project.
Can the same model move to the perimeter later?
Open models — yes: the same open-weight model deploys on your hardware with no change to application code. Flagship models are discussed separately.
How to start?
Send the scenario and expected request volume. We return model selection, limits and rates. For flagship models, we return delivery terms.
Describe the scenario — we return model selection and rates
Send what the model should do and the expected request volume. We return model selection, limits and terms. If the perimeter fits better, we say so in the reply.
batony@agent.qq.com