Models

EU GPT runs a curated set of open-weight large language models, hosted on European infrastructure. Model selection on the API is automatic — the router always picks. Pass "model": "auto"; it is the default and the only meaningful value.

Auto routing (always on)#

The router selects the best available model based on:

• Whether the request needs deep reasoning, retrieval, or fast turnaround.
• The length and complexity of the input.
• Operational cost when comparing similarly-capable models.

Why automatic selection is a win:

• Smart selection. The router routinely outperforms hand-picked model IDs on mixed workloads — task classification is good enough that hard-coding a single model would leave quality on the table.
• Free upgrades. When we add a newer, stronger model to the catalog, every caller benefits the same day without a code change.
• Resilience. If a backend is overloaded or briefly unavailable, the router routes around it.

The actual model used is reported in the response:

{
  "model": "gpt-oss-120b",
  "output_text": "…"
}

Or, in the stream, via the response.routing.completed event (when available).

Versioning policy#

• Model selection is always automatic. The auto contract is “pick the best available model”, and that contract is stable — but it has never been, and is not, a promise about which specific model you receive.
• The set of models the router may select from changes as providers’ lifecycles evolve. These changes do not affect auto’s contract.

Migration notices appear in the Changelog.