[Bug]: Default Gemini model (gemini-2.5-pro) is unavailable to new API keys, breaking every AI feature on a fresh install

[And3m]

Surface

Local main branch

---

What happened?

On a fresh clone with a newly created Gemini API key, every AI feature fails with HTTP 502. The cause is that PROVIDER_CATALOG ships gemini-2.5-pro as the Gemini default (backend/app/llm/catalog.py:17), and Google has since closed the entire Gemini 2.5 family to newly created API keys. The provider returns 404 NOT_FOUND with the message "This model models/gemini-2.5-pro is no longer available to new users." Because the default model is what a first-time user gets before touching settings, the practical effect is that graph generation, closure quizzes, and the assistant are all dead on arrival for any new user — the exact first-run path the quickstart walks people through.

This is not a quota or billing problem. The key authenticates fine and GET /v1beta/models lists both gemini-2.5-pro and gemini-2.5-flash; they only fail at generateContent. Listing a model is not the same as having access to it, which makes this easy to misdiagnose.

The same stale list is duplicated in the frontend at frontend/src/components/SettingsModal.tsx:11-14, and SettingsModal.tsx:216 selects nextOptions[0] when the provider dropdown changes — so a user who opens Settings and re-selects "gemini" is auto-assigned gemini-2.5-pro, i.e. steered straight back into the broken model.

Steps to reproduce

1. Create a new Gemini API key at https://aistudio.google.com/apikey (an older key will not reproduce this — existing users retain 2.5 access).
2. git clone, cp .env.example .env, set KG_GEMINI_API_KEY=..., start the stack.
3. Open the frontend, leave the default model untouched (gemini-2.5-pro).
4. Open the starter mathematics-demo graph, select Arithmetic and number sense, and start a closure quiz.
5. Observe HTTP 502 and "closure quiz generation failed: unexpected provider error". Any other AI action (expand graph, ingest topics, assistant) fails the same way.

Expected behavior

A fresh install with a valid, newly created provider key should be able to generate on the default model without the user first discovering that the default is unusable.

Concretely, I'd suggest:

1. Move the Gemini default_model to a model still available to new keys — gemini-3-flash-preview is already in model_options and works.
2. Refresh model_options to drop or de-prioritise the 2.5 entries.
3. Ideally, surface the provider's own message. The 404 text is fully actionable, but the user only sees "unexpected provider error" (backend/app/services/quiz_service.py:261) unless they enable debug mode and read logs/logs.log. Given the project's "fail-closed beats fake success" principle, this fails closed correctly but doesn't say why.

I'm happy to open a PR for (1) and (2) — and (3) separately if you want it scoped apart.

Logs, screenshots, or traces

Direct provider check, confirming this is availability and not quota:

POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-pro:generateContent
404 {
  "error": {
    "code": 404,
    "message": "This model models/gemini-2.5-pro is no longer available to new users. Please update your code to use a newer model for the latest features and improvements. We recommend you to use the Interactions API (https://ai.google.dev/gemini-api/docs/migrate-to-interactions).",
    "status": "NOT_FOUND"
  }
}

gemini-2.5-flash and gemini-2.5-flash-lite return the identical error. What I measured on a new key:

Model	generateContent
gemini-2.5-pro	404 — not available to new users
gemini-2.5-flash	404 — not available to new users
gemini-pro-latest	429 — quota
gemini-3-pro-preview	429 — quota
gemini-flash-latest	503 — overloaded
gemini-3-flash-preview	200 OK

From logs/logs.log on the failing quiz:

{"kind": "server", "level": "error",
 "title": "POST /api/v1/graphs/mathematics-demo/topics/arithmetics/quiz/start",
 "message": "Closure quiz generation failed", "status_code": 502,
 "response_excerpt": "{\"detail\":\"closure quiz generation failed: unexpected provider error\",\"diagnostics\":{\"model\":\"gemini-2.5-pro\",\"error_type\":\"ClientError\",\"error_message\":\"404 NOT_FOUND ... no longer available to new users\"}}"}

Confirming the fix — after POST /api/v1/workspace/config {"default_model":"gemini-3-flash-preview"}, the same quiz call returns 200 with 12 well-formed questions and "generator": "gemini-3-flash-preview". No code change was needed, which is why I think this is just a catalog default that reality moved out from under.

Commit or branch

main @ 41ad232

Regression?

Yes

[miuuyy]

Fixed in #3: replaced the default Gemini model "gemini-2.5-pro" with "gemini-3.7-flash" and updated the frontend model ordering so a fresh install and provider re-selection use the new default.