Skip to content

Adding or updating a model

Most new models need no code change. genimg resolves --model in this order:

  1. A curated alias in registry.py, such as gdm:nb2.
  2. A model id already in the registry, such as gpt-image-2.
  3. Any id a provider recognises by shape: gpt-image-* goes to OpenAI, gemini-*image* to Google.

So a new model works straight away with its full id:

genimg "a robot" --model gemini-4.0-flash-image

When to edit the registry

Add an entry only for a short alias, a quality_rank so it sorts well in genimg models, or a fixed Vertex region. Point the alias at the stable id that succeeds in a real generation. Provider listings can keep advertising a -preview id after it stops serving.

Steps

  1. Add the entry to _REGISTRY in src/genimg/registry.py:

    "gdm:<short>": ModelSpec("google", "<model-id>", region="global", quality_rank=8),

  2. Add a price if you want cost estimates. Google prices live in _PER_IMAGE in providers/google.py. GPT Image 2 and 2.5 use token grids in _TOKEN_GRID in providers/openai.py, copied from OpenAI's cost calculator; older OpenAI models use _LEGACY_PER_IMAGE. An unpriced model shows unknown.

  3. If the model takes different sizes or qualities, update the provider's capabilities(), and quality_options() for OpenAI. The CLI and Draw Studio both read them.
  4. Add a resolve test in tests/test_registry.py, and a price test in tests/test_cost.py.

Finding new models

Check the provider pages: Google models and pricing, OpenAI models and pricing.

To list image models your credentials can see that the registry lacks, run this with uv run python -:

from genimg.auth import google, openai
from genimg.providers.listing import listed_model_ids
from genimg.registry import all_canonical

known = {s.model_id for s in all_canonical().values()}
for name, client in [("openai", openai.get_client()), ("google", google.get_client())]:
    listed = listed_model_ids(client.models.list())
    new = {m for m in listed if "image" in m and "/" not in m} - known
    print(name, sorted(new))

Output on 25 September 2026, with Azure OpenAI and a Gemini API key:

openai ['gpt-image-1-2025-04-15', 'gpt-image-1-mini-2025-10-06', 'gpt-image-1.5-2025-12-16', 'gpt-image-2-2026-04-21', 'gpt-image-2.5-flare-2026-09-08', 'gpt-image-2.5-sunburst-2026-09-08']
google ['gemini-3-pro-image-preview', 'gemini-3.1-flash-image-preview']

Dated snapshots and -preview ids like these already resolve by shape, so they need no entry. On Vertex, the listing can omit Model Garden models.