Hello, I am troubleshooting intermittent Gemini API failures observed on 30 September and 1 October 2026.
Environment: Windows, Python/httpx, Gemini Developer API free tier, with a small local MCP adapter. The problem also occurs in a direct REST request outside the adapter. No automatic retry loop or paid billing was enabled.
Observed results:
- GET /v1beta/models succeeds.
- POST /v1beta/interactions with gemini-3.7-flash and gemini-3.8-flash intermittently returns HTTP 503.
- A short gemini-3.7-flash Interactions test with the same key completed successfully with the expected CONNECTION_OK response.
- Ordinary harmless text requests still returned 503.
- An independent POST /v1beta/models/gemini-3.7-flash:generateContent returned HTTP 503 with error status UNAVAILABLE. This was an explicit HTTP response, not a client timeout.
- On 30 September, eight generation attempts produced three successes, four 503 responses and one 429. Testing stopped after the 429. These counts describe that day’s small test only, not a service-wide failure rate.
Example successful short Interactions body:
{
"model": "gemini-3.7-flash",
"input": "Diagnostic connectivity check. Reply with exactly: CONNECTION_OK",
"store": false,
"generation_config": {"max_output_tokens": 128}
}
An example ordinary question that failed was: “What is the difference between a downloaded copy and an offline-synced folder in Google Drive? Answer in Russian in three brief sentences.”
Interactions requests used a 120-second timeout (15-second connect timeout). The independent generateContent test used 45 seconds (15-second connect timeout). Local mocked transport tests pass, but that does not establish live provider reliability. Successful model listing alone also does not prove that every generation request is authorized or available.
Could the team advise:
- Is there a known availability issue for these models/endpoints, including free-tier traffic?
- Which non-secret response fields or request identifiers should be captured to distinguish temporary availability problems from project/quota issues? I will not post API keys, account identifiers or raw machine logs publicly.
- What bounded retry/backoff and recovery check do you recommend, without rotating keys, creating replacement accounts, or enabling paid billing?
I am not assuming a global outage or a confirmed root cause. I would appreciate a supported diagnostic path. Thank you.