Gemini API 429 RESOURCE_EXHAUSTED: why your limit is 0

Every Gemini quota error looks the same at first: HTTP 429, "status": "RESOURCE_EXHAUSTED", “You exceeded your current quota”. The part that matters is one line further down, in the * Quota exceeded for metric lines: the limit number and the model. With a free AI Studio key I hit three different ones on purpose, and each one needs a different fix:

The details array names the quota exactly. The per-minute one was GenerateRequestsPerMinutePerProjectPerModel-FreeTier, the daily one GenerateRequestsPerDayPerProjectPerModel-FreeTier.

limit: 0: the model isn’t in your tier

This was the first call I made with the key. It wasn’t a burst: a single request to gemini-3.1-pro-preview came back 429 straight away. Same for gemini-pro-latest. On the free tier these models have a quota of zero. The error still says “Please retry in 10h17m”, but that only marks when the daily counters reset. Retrying at that time gets the same limit: 0.

const res = await ai.models.generateContent({
  model: 'gemini-3.5-flash',
  contents: 'hi',
});

Per-minute limit: wait the time it gives you

Sending one tiny request after another to gemini-3.5-flash, calls 1 to 7 succeeded, and call 8 failed 8.4 seconds in:

* Quota exceeded for metric: generativelanguage.googleapis.com/generate_content_free_tier_requests, limit: 5, model: gemini-3.5-flash
Please retry in 46.130728812s.

The limit says 5, but seven got through first, so don’t use the number to pace yourself exactly. The @google/genai SDK (2.27.0) threw this straight to my code: it didn’t retry on its own.

import { GoogleGenAI, ApiError } from '@google/genai';
const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });

async function generateWithRetry(params, maxWaits = 4) {
  for (let attempt = 0; ; attempt++) {
    try {
      return await ai.models.generateContent(params);
    } catch (err) {
      if (!(err instanceof ApiError) || attempt >= maxWaits) throw err;
      if (err.status === 503) { // "high demand": back off and try again
        await new Promise((r) => setTimeout(r, 2000 * 2 ** attempt));
        continue;
      }
      if (err.status !== 429) throw err;
      const body = JSON.parse(err.message).error;
      if (/limit: 0\b/.test(body.message)) throw err; // no quota for this model at all
      const delay = body.details?.find((d) => d['@type']?.endsWith('RetryInfo'))?.retryDelay ?? '10s';
      const ms = Math.ceil(parseFloat(delay) * 1000) + 500;
      if (ms > 120_000) throw err; // a daily quota: the delay runs to midnight UTC
      await new Promise((r) => setTimeout(r, ms));
    }
  }
}

Daily limit: it resets at midnight UTC, not Pacific time

My test calls used up gemini-3.5-flash’s free daily quota (limit: 20). The retry delay pointed at 2026-10-08 00:00:13 UTC. That’s 8 PM in New York and 5 PM in San Francisco, not midnight Pacific as a lot of older answers say.

Not a 429: “This model is currently experiencing high demand”

While I was testing, gemini-3.8-flash (the newest Flash) returned this on every call, and later gemini-3.5-flash did too, for a few calls:

{
  "error": {
    "code": 503,
    "message": "This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.",
    "status": "UNAVAILABLE"
  }
}

This is Google being overloaded, not your quota. There’s no retry delay in it, and other models kept working at the same moment. Calls that did go through during that period were slow, too: one took 56 seconds.

What didn’t work

How this was tested

A free Google AI Studio API key on October 7, 2026, with @google/genai 2.27.0 on Node 24.14.0 (Windows 11). Every request was one word with maxOutputTokens: 1, sent one after another, never in parallel. Limits were reached on purpose: a Pro model for limit: 0, a fast loop for the per-minute limit, and the day’s test calls for the daily one. The limits shown are the free tier’s on that day. Google changes them, so read the numbers in your own error message. I didn’t test paid tiers, Vertex AI, or the Gemini CLI, whose 429s come from a different quota.

The code for every case above is public, so you can run it yourself: gemini-429 in nk-repro.