Models & capabilities
What a trustworthy model listing must describe before it can receive buyer traffic.
In this chapter
The catalog today#
No verified model supply is published by the current customer API. /api/v1/models therefore returns an empty list. The workspace does not populate that list with famous model names to make the product appear complete.
A model available through an API is not the same as a subscription to the model provider's consumer application. Access to a GPT model, for example, would not automatically include every feature of a ChatGPT subscription.
The model listing contract#
| Field | Why it matters |
|---|---|
| Exact model and version | Buyers must know what they are requesting, including relevant quantization or fine-tuning differences. |
| Input/output modalities | Text, images, audio, and video are different serving capabilities. |
| Context and output limits | A low price is not useful if the workload does not fit. |
| Supported API features | Streaming, tools, structured output, and other options must be tested, not assumed. |
| Input/output rates and fees | A single unexplained token price can hide materially different costs. |
| Serving region and privacy | Where data goes and which handling commitments apply. |
| Measured service quality | Latency, throughput, successful completion, and observed availability. |
Compare equivalent workloads#
Supplier competition only helps if the offered services are comparable. Model version, quantization, prompt length, generated length, caching, batching, latency targets, and privacy constraints can all change the comparison. A smaller model is not automatically a discounted equivalent of a more capable proprietary model.
A listing should distinguish supplier declarations, platform measurements, contractual commitments, and hardware-backed evidence. An ‘online’ status or successful test question alone cannot prove the exact weights being served.
When model access can open#
- Obtain an authorized serving endpoint and complete its security and ownership checks.
- Confirm model licensing and permitted downstream use.
- Benchmark the endpoint under representative workloads and failures.
- Establish pricing, usage metering, and cost limits.
- Publish accurate capabilities and data-handling information.
- Enable bounded buyer access only after accounting and operational controls pass.