google_imagen: all three hardcoded imagen-4.0 models return 404 'no longer available to new users'

[srikanthapril19-code]

Summary

google_imagen offers exactly three models, and all three are now closed to new Google accounts. Any user who signs up today gets a 404 no matter which they pick, so the tool is unusable on a fresh install — while the registry still reports it as available, because availability is gated on GOOGLE_API_KEY being present rather than on the model being reachable.

The model ids are hardcoded in tools/graphics/google_imagen.py — an enum of three at L118-120 and the same default repeated at L122, L177 and L215.

Operating system

Windows 11, Python 3.12

Pipeline

Any pipeline selecting Google Imagen for stills

Runtime / renderer

n/a — fails at the API call

Steps to reproduce

1. Create a new Google API key (AI Studio or Cloud Console) — this matters; existing accounts are grandfathered and will not reproduce it.

2. Set GOOGLE_API_KEY in .env.

3. Run google_imagen:

   registry.get("google_imagen").execute({
       "prompt": "a plain grey studio backdrop",
       "output_path": "out.png",
   })

4. Repeat with model set to each of the other two enum values.

Expected behavior

An image is generated, or the tool reports itself unavailable with actionable guidance.

Actual behavior

All three models return 404. The tool reports available beforehand, so the failure only surfaces at call time, mid-pipeline.

Note the models are still listed by the API and still advertise predict, so a catalog check does not reveal the problem — only an actual call does:

models/imagen-4.0-generate-001        ['predict']
models/imagen-4.0-ultra-generate-001  ['predict']
models/imagen-4.0-fast-generate-001   ['predict']

Relevant logs or error output

imagen-4.0-generate-001        HTTP 404: This model models/imagen-4.0-generate-001 is no longer available to new users.
                                        Please update your code to use a newer model for the latest features and improvements.
imagen-4.0-fast-generate-001   HTTP 404: This model models/imagen-4.0-fast-generate-001 is no longer available to new users.
imagen-4.0-ultra-generate-001  HTTP 404: This model models/imagen-4.0-ultra-generate-001 is no longer available to new users.

---

Suggested fix

The current-generation image models on the Generative Language API are reached through :generateContent, not :predict, and return the image as an inlineData part:

gemini-3.1-flash-image
gemini-3.1-flash-image-preview
gemini-3-pro-image
gemini-3-pro-image-preview
gemini-2.5-flash-image
nano-banana-pro-preview

I could not verify these end-to-end — my account returned 429 You exceeded your current quota for image generation on the free tier, so I can confirm the old path is dead but not that the new path works. Flagging that explicitly rather than implying I tested it. Happy to verify if someone with billing enabled can confirm, or once I enable it.

Two things worth considering alongside the model list:

1. Don't hardcode the default in four places. L122, L177 and L215 each repeat "imagen-4.0-generate-001", so a fix has to touch all of them consistently.
2. get_status() currently only checks that GOOGLE_API_KEY is set. That is why this reports available and then fails. This same pattern bit me elsewhere in the repo — see .env.example inline comments load as credential values, making ~30 tools falsely report available #431, where empty keys pick up their own .env.example comment as a value and ~30 tools report available with garbage credentials. A status check that reflects reachability, not just key presence, would catch both classes.

Related: #431.