Hello,
I am testing the Gemini Developer API from Google Apps Script using UrlFetchApp (direct REST, no SDK). My Google AI Studio project is on the Free tier with no billing enabled. I am trying to evaluate an article-review prototype, but no full article review has succeeded.
Could you help determine whether these repeated 503 responses reflect current serving capacity or a project/model-specific availability issue, and advise what additional non-sensitive diagnostics would be useful?
Environment
- Gemini Developer API, not Vertex AI or the Gemini consumer app.
- Endpoint:
POST https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent - Models tested:
gemini-3.8-flashandgemini-3.5-flash-lite. - Authentication: API key in the
x-goog-api-keyheader. The key is not included here. - Content-Type:
application/json. - Text-only requests; no files, tools, grounding or cached content.
- Requests sent individually. No parallel calls, automatic retries or automatic model fallback in this prototype. Calls returning 5xx trigger a local five-minute cooldown.
Observations on September 28-29, 2026
- A short JSON diagnostic on
gemini-3.8-flashsucceeded at2026-09-28T13:20:31.471Z(22:20 JST): HTTP 200, 11 input tokens, 5 output tokens, with the requested JSON returned. - Article-review requests failed with HTTP 503 /
UNAVAILABLEon both models. A Flash-Lite request containing 13,024 article HTML characters was 21,994 bytes and failed in 602 ms. - Truncating only the article HTML to 2,000 characters reduced the payload to 6,506 bytes, but still returned 503 in 618 ms on Flash-Lite.
- A fixed short diagnostic, with no article or system instruction, also failed on Flash-Lite at
2026-09-28T15:19:24.024Z(00:19 JST on September 29): 230-byte payload, output limit 256, HTTP 503 /UNAVAILABLE, 316 ms. - A subsequent one-shot diagnostic kept the same Flash-Lite model and short prompt but omitted
responseMimeType,thinkingConfigandtemperature. OnlymaxOutputTokens: 256remained in generationConfig. It also returned HTTP 503 /UNAVAILABLE.
Our error logger detected capacity/high-demand/overload wording in these responses. It does not retain the complete raw error message, so this description is not a verbatim quotation. We have not captured a Google-generated request ID; our own diagnostic UUIDs are not server request IDs.
Minimal-settings request that also failed
{
"contents": [
{
"role": "user",
"parts": [{"text": "Return exactly this JSON: {\"ok\":true}."}]
}
],
"generationConfig": {"maxOutputTokens": 256}
}
This was sent to the endpoint above with gemini-3.5-flash-lite. The error occurs without article content, a system instruction, a large output allowance, structured-output configuration or an explicit thinking setting. This does not by itself establish the root cause.
Usage and status checked
AI Studio showed Free tier for the intended project. Its usage dashboard recorded the API requests and 503 errors. The rate-limit page displayed these peak usage/limit figures for the selected 28-day period:
| Model | RPM | TPM | RPD |
|---|---|---|---|
| Gemini 3.8 Flash | 1 / 5 | 5.18K / 250K | 5 / 20 |
| Gemini 3.5 Flash Lite | 1 / 15 | 5.18K / 250K | 3 / 500 |
These are dashboard snapshots, not real-time remaining quotas; the page notes that metrics can lag by up to 15 minutes. None of the displayed thresholds was exhausted at the time checked.
The public AI Studio status page showed All Systems Operational around 00:20 JST on September 29. I understand that this does not exclude partial or project-specific issues. The Logs and Datasets page required billing setup to enable detailed logging; I have not enabled billing.
Questions
- Are there known serving-capacity or availability issues affecting these models or newly created Free-tier projects during the timestamps above?
- Is any aspect of the minimal request or the Apps Script execution environment relevant to these 503 responses?
- Apart from bounded backoff/retry, what evidence should I collect to help investigate? If a project identifier is needed, please provide an official private channel where it can be shared safely.
I would like to keep this evaluation on the Free tier. I am not requesting a quota increase or assuming that enabling billing would resolve the issue.
Thank you.