A Gemini 429 or RESOURCE_EXHAUSTED is a signal to diagnose the failure, not an instruction to retry forever. First distinguish a fixed quota or spend limit from temporary capacity pressure; then apply bounded retries, reduce avoidable demand, and choose a fallback that fits the request’s latency and quality requirements. Gemini API and Vertex AI have different error guidance and retry recommendations, so keep their policies separate.
What does a Gemini 429 or RESOURCE_EXHAUSTED error mean?
The status alone may not tell you whether another attempt can succeed. The likely cause depends on which Google service surface you are calling and on the error details.
Gemini API: identify the specific limit or service error
Gemini API limits can cover requests per minute, input tokens per minute, requests per day, model-specific dimensions, and, for eligible accounts, spend over a rolling ten-minute window. Limits are applied at the project level, not separately to each API key. Current values vary by model, tier, and account status, and Google cautions that published limits do not guarantee capacity. Rotating API keys therefore does not increase a project’s quota.
Google AI for Developers’ Gemini API error reference distinguishes rate_limit_exceeded and too_many_requests, which indicate short-term rate or burst limits, from quota_exceeded, which indicates a daily quota. It maps temporary service overload or downtime to HTTP 503 service_unavailable. Check the returned error details and the project’s live limits before changing your retry or routing policy.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Where spend-based limits apply, the Gemini API Rate Limits page lists Tier 1, Tier 2, and Tier 3 limits of $10, $50, and $200, respectively, per rolling ten-minute window. These are tier-dependent vendor-published figures, not universal account allowances; check the live limits for your project.
Vertex AI: quota exhaustion and shared-server overload can both produce 429
Google Cloud’s Vertex AI API Errors guidance says HTTP 429 RESOURCE_EXHAUSTED can mean either that a quota was exceeded or that shared-server capacity is overloaded. Inspect the error message and the relevant quota. A retry may help with temporary overload, but it cannot permanently fix a fixed project quota. Correct invalid requests rather than retrying them.
Do not transfer an error interpretation or retry policy from Gemini API to Vertex AI just because both can return 429. The services have separate guidance, quota surfaces, and capacity options.
How should you decide whether to retry, wait, or fail?
Classify the failure before scheduling another attempt. A short-term rate limit or temporary service error may clear; a daily quota or spend cap usually requires waiting for its reset window, reducing demand, or changing the relevant limit or capacity arrangement. A malformed request, authentication failure, permission denial, or billing issue requires correction, not repetition.
Recommended Free Tools
- Potentially transient: Gemini API 429 rate or burst errors and 503 service-unavailable errors; Vertex AI 429s when the details point to shared-server overload; and other transient statuses supported by your client’s policy.
- Not a retry fix: a daily quota or fixed spend cap, invalid input, missing credentials or permissions, and billing problems.
- Unclear: retain the status and full error details, inspect the project’s applicable quota, and avoid an aggressive retry loop until the cause is better understood.
Google’s Gemini API troubleshooting guidance recommends exponential backoff for retryable errors such as 429 and 503, and warns against treating 400, 402, or 403 as transient. For direct REST calls or custom retry logic, add random jitter, limit attempts, and set an overall deadline. Jitter spreads retries across time so clients that failed together do not all send another request at once.
How many times should you retry Gemini requests?
Use the guidance for the specific API surface and client in production. The following are Google-published recommendations or defaults, not a guarantee that a request will succeed:
| Surface or client | Documented retry guidance | Important qualification |
|---|---|---|
| Gemini API Python SDK | Up to four retries for transient errors, with an initial delay of approximately one second and a maximum delay of 60 seconds. | These are defaults reported by Google AI for Developers’ Gemini API Troubleshooting guide, accessed in 2026. Verify behavior against the deployed SDK version because defaults can change. |
| Vertex AI | No more than two retries, with a minimum initial delay of one second and subsequent requests backing off exponentially. | This is the platform-specific guidance in Google Cloud’s Gemini Enterprise Agent Platform API Errors documentation, last updated October 1, 2026. |
For custom logic, cap both the number of attempts and total elapsed time. A retry that starts after the caller’s deadline cannot help the user. Preserve idempotency where it matters, and record status, error details, attempt count, and elapsed time so that repeated failures can be diagnosed.
Make sure retries are bounded across the whole request path. If the SDK, application, queue, and gateway each retry independently, their attempts can multiply into a retry storm. Assign a clear retry owner or coordinate the limits across layers, and avoid immediate retries: Google Cloud’s “Reduce 429 errors on Vertex AI” specifically says, “An immediate retry is not recommended.”
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallHow can you prevent overload before adding a fallback?
Fallback is only one resilience layer. Reducing unnecessary work and smoothing arrivals can prevent avoidable limit pressure and make any recovery path more useful.
Reduce repeated token work
- Shorten prompts and specify only the output length and format the task needs.
- Summarize or trim long context where that preserves the information needed to answer.
- Cache repeated context where appropriate instead of sending it again for every request.
Smooth traffic instead of releasing bursts
Use admission control, queues, or rate limiting to spread work rather than allowing synchronized jobs or client retries to arrive in a sharp burst. This is especially important after a service recovers: releasing a backlog all at once can recreate the pressure that caused requests to fail.
Match capacity and routing to the workload
Google Cloud’s Vertex AI guidance recommends considering the global endpoint where appropriate, so requests can be routed across regions rather than relying only on one regional endpoint. Whether that fits depends on the workload and deployment requirements; check the current endpoint and model availability for your use case.
The same guidance differentiates service tiers by traffic shape: Priority PayGo for critical, unpredictable user-facing traffic; Provisioned Throughput for consistently high real-time traffic; and Flex or Batch for latency-tolerant or asynchronous work. Confirm current product terms, model availability, and suitability before adopting a tier. Circuit breaking and graceful failure at the gateway can also contain failures; Google’s guidance names Apigee as one option.
Which fallback pattern fits the failure?
Choose the response by failure type, recovery time, and the caller’s tolerance for delay. Do not treat a switch to another model or provider as an automatic remedy for a project quota, client error, or billing problem.
| Situation | Suitable response | Trade-off to account for |
|---|---|---|
| Short-lived rate pressure or temporary service overload | Retry the same request with exponential backoff and jitter, within an attempt cap and deadline. | Consumes time and may add cost; stop when the retry or latency budget expires. |
| Work can wait longer than an interactive request | Queue it for later processing, or use an appropriate asynchronous or batch path. | Results arrive later; the application needs a way to track pending work and communicate its status. |
| Interactive response is required, but the primary path remains unavailable | Return a deliberate degraded response, such as asking the user to try again, or route to a prevalidated alternative. | A degraded response may provide less functionality. An alternative model or provider can change output quality, behavior, privacy implications, and cost. |
| Persistent high-volume real-time demand | Evaluate capacity options such as Provisioned Throughput against the traffic pattern and service requirements. | Capacity arrangements have their own availability, model, and commercial terms; verify current details with Google Cloud. |
| Fixed quota or spend limit reached | Reduce or defer demand, verify the limit and reset conditions, or review the project’s applicable capacity and limits. | A fast retry or model switch within the same constrained project may not address the limiting condition. |
A useful application design sets a retry budget and a latency budget. Retry only while both permit another attempt. When either expires, apply the planned degraded response or route to an independently available alternative. This is an application-level design choice, not a universal Google-prescribed fallback sequence.
What should you validate before switching models or providers?
A fallback is safe only if the application can tolerate what changes. Test it against representative requests before enabling automatic routing, and define which failures are eligible to trigger it.
- Output contract: confirm that the alternative produces the required schema, fields, and formatting, including how malformed or incomplete output is handled.
- Tools and safety: verify tool calling, safety behavior, and any application assumptions about model responses; do not assume they match the primary model.
- Quality and latency: establish whether the alternative is acceptable for the task and whether it can respond within the remaining deadline.
- Privacy and data terms: check what data is sent to the alternate service and whether that path complies with the application’s requirements.
- Cost and retry behavior: account for the cost of the original attempt plus retries and fallback work, and ensure the new path cannot trigger another unbounded retry chain.
Google’s cited guidance covers retrying, quotas, routing, capacity tiers, and gateway-level resilience; it does not prescribe a universal cross-provider cascade. The application owner must choose and validate that behavior.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteQuick Recap
Operational checklist for a controlled recovery path
- Capture the API surface, HTTP status, exact error name and details, model, project, and time of failure.
- Check the relevant live project quota or spend limit; do not assume API keys have separate allowances.
- Classify the error as transient, quota-related, or a client/account issue before retrying.
- Apply surface-appropriate retries with exponential backoff, jitter, an attempt cap, and a request deadline.
- Reduce unnecessary tokens and smooth incoming work; avoid releasing retries or queued work as a burst.
- When the retry or latency budget ends, use the defined degraded response or a tested alternative path.
- Monitor errors, retries, latency, quota pressure, and fallback use so persistent capacity or demand problems are addressed rather than hidden by retries.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
