RESOURCE_EXHAUSTED on the Gemini API: which of the two 429s you have

Two error formats, because there are now two APIs

Before diagnosing anything, work out which surface you are talking to, because the error envelopes differ and only one of them contains the string RESOURCE_EXHAUSTED.

The current Interactions API

Google recommends this one now, and it uses a flatter envelope with a snake_case string in code and no status field at all.

{
  "error": {
    "code": "rate_limit_exceeded",
    "message": "..."
  }
}

The legacy generativelanguage endpoint

Still live, still widely used, and still returning the classic google.rpc.Status envelope with a numeric code and a screaming-case status. This is where the searched-for term comes from.

{
  "error": {
    "code": 429,
    "message": "...",
    "status": "RESOURCE_EXHAUSTED",
    "details": [ /* may carry QuotaFailure or RetryInfo */ ]
  }
}

rate_limit_exceeded versus quota_exceeded

This split is the most useful thing on the page, because the two conditions need opposite handling and both arrive as a 429.

A retry loop that cannot tell them apart will spend its whole budget of attempts against a daily quota that has fourteen hours left to run, and then report a timeout. Branch first.

def handle_429(err):
    code = err.code if isinstance(err.code, str) else None

    if code == "quota_exceeded":
        # RPD resets at midnight Pacific. Retrying now cannot work.
        return failover_or_defer()

    if code == "rate_limit_exceeded" or err.status == "RESOURCE_EXHAUSTED":
        return retry_with_backoff()   # seconds, not hours

    raise err

The limits Google stopped publishing

This section exists because almost every other page on this topic is now wrong, and we would rather tell you what is unknowable than repeat a stale number.

That is the whole of it. The per-model RPM, TPM, and RPD tables that used to sit on that page, the ones every tutorial quotes as "15 RPM and 1,500 RPD on the free tier", have been removed. There is no citable public number any more. Open AI Studio and read your own project's limits; that is now the only accurate source.

What the page does still state, and what therefore remains safe to build on:

• "Rate limits are applied per project, not per API key." Minting a second key changes nothing. This one surprises people every time.
• "Requests per day (RPD) quotas reset at midnight Pacific time." Not at midnight your time, and not on a rolling 24-hour window.
• Limits are per model, not shared. Being throttled on one model says nothing about another, which makes a same-provider fallback genuinely useful here.
• Preview and experimental models are more restricted than GA models, explicitly.
• Image models are limited in IPM (images per minute), and some models carry a TPD as well as a TPM.
• Priority inference gets 0.3x the standard rate limit for its model and tier. You trade throughput headroom for latency.

Note that unlike OpenAI, Google kept the waiting period: Tier 2 and Tier 3 both require elapsed days as well as spend, and the spend counted is cumulative across Google Cloud services on the linked billing account, not just the Gemini API. Free to Tier 1 takes effect essentially instantly; later upgrades land within about ten minutes.

The free tier and what it costs you

The Gemini free tier is genuinely generous and it has one condition that belongs in any honest discussion of it. From Google's own pricing tables, every model has a row labelled "Used to improve our products": Yes on the free tier, No on paid.

The practical upgrade path is short: link a billing account and you are Tier 1, which raises your limits, removes the training clause, and does not oblige you to spend anything beyond what you use.

Fixing it

Google, unlike Anthropic, documents jitter explicitly, and its own SDK already implements a reasonable policy: "the Python SDK automatically retries transient errors up to four times with an initial delay of approximately 1 second and a maximum delay of 60 seconds."

One last structural note for anyone comparing the two Google surfaces. Vertex, now branded Gemini Enterprise Agent Platform, does not work like this at all: it allocates org-level tokens-per-minute by 30-day rolling spend and has no per-tier RPM limit, bursting above baseline on a best-effort basis. Its own documentation warns that "high and instantaneous traffic can lead to throttling even if your average per-minute usage is below your limit", which is the same acceleration-limit idea both OpenAI and Anthropic have under different names.

Questions people ask

Keep reading

Continue through Errors and fixes with The model does not exist or you do not have access to it.