The three names
This is the part worth reading carefully, because a model has three names and they do different jobs.
The gateway name is built from the provider name and an alias you choose. If
you leave the API name blank, the alias is sent upstream as well.
Set the API name when the provider’s own identifier is something you would
rather not expose. A model that upstream calls
claude-fable-5-1 can be anthropic/fable-5.1 to everything you run.
The gateway name is the only name accepted from a client. The provider’s own
identifier does not work from outside, so nothing you run can reach a model
you did not deliberately expose.
Automatic models
Adding a model to an automatic provider means picking it from the catalog. Everything factual about it, the context window, the costs, whether it can see images or call tools, is read live from models.dev on every page load. You can still rename it. The edit form’s gateway name field changes the public name without touching what is sent upstream. If a model leaves the catalog or is deprecated, it is disabled automatically and labelled, so it stops counting against your plan once you delete it.Custom models
For a custom provider you describe the model yourself, across three tabs. General is the names, the type and the display name. Types are chat, embedding, moderation, vision, audio and image. Pricing is costs per million tokens for input and output, optional cache read and write costs, an optional higher rate above a context threshold, and per image or per minute costs for those model types. Also the context window and the maximum output. Configuration is what the model can do: attachments, reasoning and tool calling, plus how reasoning and speed should be expressed upstream.Reasoning style
Different providers ask for reasoning effort in different ways. Left on automatic, nemu infers it from the model. Set it explicitly when the guess is wrong.
Effort runs low, medium, high, xhigh and max, clamped to what each model
actually supports.
Speed mode does the same for the fast toggle, either Anthropic’s fast mode
or OpenAI’s priority service tier.
Two names that always work
auto asks the gateway to choose an enabled chat or vision model. It ranks them
by the cost stored against each one, but a model added from the models.dev
catalog stores no cost, because its pricing is resolved live at request time.
Those all tie, so the choice usually falls to the first provider and model name
alphabetically. Read auto as “some enabled model” rather than “the cheapest
one”.
To decide what it means yourself, build a workflow whose trigger is
auto. The gateway’s own choice is used only when no workflow claims it.
A suffix in square brackets is a variant. anthropic/fable-5.1[1m] addresses
the same model with a one million token context window.
The window is only ever raised, never lowered, so the suffix is redundant on a
model that already reaches a million tokens and is accepted there anyway. A
client can send it on every model name without tracking which ones need it.
[1m], [1m-context] and [context-1m] are the same variant, and case and
spacing inside the brackets do not matter. Any other variant is ignored rather
than refused. The gateway covers the detail.