Add gateway to Known Model Names

[dsfaccini]

No description provided.

[dsfaccini]

there's a flaky outlines test breaking CI :(

[DouweM]

I realize you couldn't see this comment in a private Slack channel, but I responded to Samuel (and he agreed):

Agreed, but if we add just a few per provider, people are definitely gonna ask us “but why not this other one?”, so maybe just include them all?

So we should NOT hard-code this, but dynamically build this based on the known model names of the providers that are known to work with gateway/{provider}

[dsfaccini]

this is now done

[DouweM]

Could we use the pydantic_ai.providers.gateway.UpstreamProvider type (possibly needs splitting up into 2 so that we don't include the APi types here), and then use those names as keys into a dict of things like {'openai': OpenAIModelName}? That way when we update it in the gateway file, this list will update too

[dsfaccini]

will check it out

[DouweM]

ApiFormats? Or do we call them API Flavors now?

[dsfaccini]

I just copied the comment as was, but yeah flavor's definitely correcter (correcter than the word correcter)

[DouweM]

This shouldn't be a public method :)

[DouweM]

Let's move this mapping to test_model_names.py so that the get_model_names used there can also take a provider name.

Then [f'grok:{n}' for n in get_model_names(GrokModelName)] could become [f'grok:{n}' for n in get_model_names('grok')], or prefixed_model_names('grok'). And then if we have the complete dict, we can basically just to model_names = [prefixed_model_names(provider) for provider in provider_model_names.keys()] or something like that.

Basically clean up all this duplication:

anthropic_names = [f'anthropic:{n}' for n in get_model_names(AnthropicModelName)]
    cohere_names = [f'cohere:{n}' for n in get_model_names(CohereModelName)]
    google_names = [f'google-gla:{n}' for n in get_model_names(GoogleModelName)] + [
        f'google-vertex:{n}' for n in get_model_names(GoogleModelName)
    ]
    grok_names = [f'grok:{n}' for n in get_model_names(GrokModelName)]
    groq_names = [f'groq:{n}' for n in get_model_names(GroqModelName)]
    moonshotai_names = [f'moonshotai:{n}' for n in get_model_names(MoonshotAIModelName)]
    mistral_names = [f'mistral:{n}' for n in get_model_names(MistralModelName)]
    openai_names = [f'openai:{n}' for n in get_model_names(OpenAIModelName)]
    bedrock_names = [f'bedrock:{n}' for n in get_model_names(BedrockModelName)]
    deepseek_names = ['deepseek:deepseek-chat', 'deepseek:deepseek-reasoner']
    gateway_names = [
        f'gateway/{provider}:{model_name}'
        for provider, model_names in gateway_provider_to_model_names().items()
        for model_name in get_model_names(model_names)
    ]
    huggingface_names = [f'huggingface:{n}' for n in get_model_names(HuggingFaceModelName)]

By having just a dict of provider names to model name types (or functions like in the case of get_heroku_model_names)