Models

List the models

Needs Call models. The answer has OpenAI’s shape:

{
  "object": "list",
  "data": [
    {
      "id": "openai/gpt-oss-120b",
      "object": "model",
      "created": 1790553600,
      "owned_by": "openai",
      "name": "GPT OSS 120B",
      "status": "available",
      "inputs": ["text"],
      "context_window": 128000,
      "max_output_tokens": 16000,
      "default_output_tokens": 4096,
      "tools": true,
      "reasoning": true,
      "units": { "input": 20000, "cached": 20000, "output": 79000 },
      "processed_in": "Germany",
      "retires_at": null,
      "replaced_by": null
    }
  ]
}

• status is preview, available, or deprecated. A deprecated model has retires_at, the date it stops answering, and replaced_by, the model to move to.
• units are the AI units per million tokens read, read from the cache, and written.
• processed_in is the country where the model runs.

The models

• Context is the most tokens the messages and the answer can hold together. A call that does not fit is refused with context_too_long.
• Default output is the answer’s limit when you leave out max_tokens.
• Reasoning models think before they answer. Their reasoning comes back in reasoning, apart from the answer in content, and counts toward output tokens. Set how much they think with reasoning_effort.
• Input units and Output units are AI units per million tokens read and written. See AI units.
• Processed in is the country where the model runs, and where what you send to it goes. Models processed in Germany keep it in the EU; the Nemotron models run in the United States.
• The Nemotron models are in preview. They use no AI units, so they are free, and your plan’s requests and tokens a minute still apply. They can be slower, briefly unavailable, or change without notice, so use them to try ideas rather than for work that must not fail.

Model terms

Each model comes with its maker’s terms, which apply to how you use it:

• GPT OSS 120B and 20B: Apache 2.0.
• Nemotron 3 Ultra, 3 Super, 3.5 Lightning, and 3 Nano Omni: the NVIDIA Open Model License.

When a model changes

A model’s name never changes. When a model is to be retired, we mark it deprecated at least 90 days before, and say which model replaces it.