Models

The catalogue

The model field of a request is a route id, exactly as printed below, punctuation included. Prices are PKR per million tokens. This table is generated from the same catalogue the dashboard renders, so it cannot drift from what the gateway actually serves.

A route marked Coming soon is listed ahead of opening and answers 404 until it does. Every route in the table has a published rate. A price shown as Not set would mean i2 has not yet published a rate for that route; it would not mean the route is free.

A route id marked proposed, not registered yet is the string i2 intends to publish for a route that has not opened. It is there so an integration can be written against the shape of it; expect to change the exact string when the route opens, and expect 400 unknown model alias if you send it before then.

Pools

• Global pass-through. Routes served through i2's upstream partners, in the OpenAI shape, with i2's keys, budgets, telemetry and rupee billing wrapped around them. Live today.
• PK NPU-native. Routes served on Ascend NPUs inside Pakistan, so prompts and completions never leave the country. Opening shortly; the routes are in the catalogue so integrations can be written against the ids now.

The pool a route lives in does not change the contract: same endpoint, same request shape, same key.

Route ids stay stable

An id names a route, not a specific checkpoint. When i2 moves the weights behind a route (a newer release of the same model, a different quantisation, a new host), the id you call stays the same and your code does not change. The request log records the resolved upstream model on every request, so you can always see what actually served a call.

Listing models over the API

GET /v1/models is not served yet; a stock OpenAI SDK's models.list() returns 404. Use this page, the dashboard's Model APIs page, or the key's allowed_models as the source of truth until the listing endpoint ships.

Which model?

• For general assistants and tool use, start with gpt-oss 120B: an open-weights reasoning model with strong published scores and a large context window.
• For the fastest first token on short prompts, pick the smallest live chat route in the table and measure; the dashboard's per-model latency and time-to-first-token charts are built for exactly this comparison.
• Pedigree (lab, parameters, licence, context window, published benchmarks) for every route is on the dashboard's Model APIs page. Those scores are the labs' own, not re-measured on i2.