Home / Blog / Pricing

Pricing

Gemini CLI 503 "The Model Is Overloaded": What It Means and Fixes

Gemini CLI error 503 'The model is overloaded'? What it means, how it differs from 429 quota errors, and what you can do now, from Google's docs and issues.

By Engineer Foodie Desk ✓ Sources checked Nov 2, 2026

8 min read

Error text checked Nov 2, 2026 on Gemini CLI’s official GitHub repo (issue #4360 and issue #7227). Error meanings and retry advice checked Nov 2, 2026 on Google’s Gemini API errors and troubleshooting pages. Model fallback checked Nov 2, 2026 in the Gemini CLI docs. Other dates under Sources. No affiliate links on this page (how we work).

Lead finding: a 503 is Google telling you its servers are busy. It isn’t a verdict on your quota, your key or your code. Google’s own error docs say to wait and retry with backoff, and a Gemini CLI maintainer said the same thing in plainer words. That’s the whole fix. It’s also why this error is so annoying.

The exact error: The model is overloaded. Please try again later.

This is how it appears in Gemini CLI’s official repo, quoted by a collaborator on issue #4360 (verified Nov 2, 2026):

[API Error: got status: UNAVAILABLE. {"error":{"code":503,"message":"The model is overloaded. Please try again later.","status":"UNAVAILABLE"}}]

One-line answer: Google’s servers didn’t have capacity for that model at that moment. It’s a server-side 503, not your quota, so waiting and retrying is the documented fix.

Google’s API errors page defines service_unavailable (HTTP 503) as “The service is temporarily overloaded or down.” Its suggested action is: “Wait and retry with exponential backoff” (verified Nov 2, 2026).

Some users have also posted a shorter variant on issue #7227: [API Error: The service is currently unavailable. (Status: UNAVAILABLE)]. It’s the same status family.

The paper trail tells you a lot. Issue #4360 opened Jul 17, 2025, collected 88 comments, and was folded into #7227 (“Increased number of 503 errors”) on Sep 16, 2025. On Jan 22, 2026 a collaborator posted: “Currently experiencing higher than usual capacity constraints. Should be temporary. We apologize for the experience you are seeing, not what we expect nor want to have users experiencing.” #7227 was then closed as not planned on May 13, 2026, with “Please reopen if still happening with latest gemini-cli version.” A third report, #12932 (“503. The model is overloaded.”), was closed by the stale bot on Feb 15, 2026. All dates verified Nov 2, 2026.

I appreciate the apology, honestly. But “should be temporary” has been the vibe for over a year of issue threads. This is Waiting for Godot with a terminal open.

503 vs 429 vs 400 vs 500: whose fault is it?

From Google’s API errors page (verified Nov 2, 2026):

CodeMeaningYour fault?
503Server overloadedNo
429Quota or rate hitYes, limits
400Bad requestUsually
500Server errorNo
  • 503 UNAVAILABLE: “temporarily overloaded or down”. Wait and retry with exponential backoff.
  • 429: either rate_limit_exceeded (per-minute or per-second limits) or quota_exceeded (“You have exceeded your daily quota”). This one is about your limits. Full breakdown in Gemini CLI 429 resource exhausted.
  • 400 invalid_request: something in the request itself is wrong. See Gemini CLI 400 invalid argument.
  • 500 api_error: “An unexpected error occurred on the server.” Google’s advice: “Retry the request. If it persists, contact support.”

Google’s troubleshooting page puts 429 and 503 in the same “retry” bucket. It says: “Only retry on transient errors (like 429, 408, or 5xx). Do not retry on client errors (like 400, 402, or 403)” (verified Nov 2, 2026). Inside Gemini CLI, a bot triage note on #7227 (May 9, 2026) says 503s are wrapped as a RetryableQuotaError in the code. It also says “the user-facing output does not clearly communicate the retry status or the specific model origin.” So if a 503 looks like a quota problem in your terminal, that’s why. Per Google’s error table, it isn’t one.

One caveat from the threads: a user on #4360 said their “overloaded” errors traced back to “the wrong api key, which was bound to free account”. That’s a user report, not Google’s diagnosis. It’s still worth checking which key or project you’re actually using.

Fixes, in order

  1. Wait, then retry. This is the documented fix. Google’s troubleshooting page recommends exponential backoff: start around 1 second, then 2, 4, 8, add jitter, and cap the number of retries. The Python SDK itself retries “up to four times” (verified Nov 2, 2026). In the CLI, that means giving it a minute rather than hammering Enter. It’s the Dumb and Dumber school of capacity planning: so you’re telling me there’s a chance. There is. Space out your attempts.
  2. Let the CLI fall back, or switch models yourself. Gemini CLI’s model routing docs say it “automatically switches to a fallback model in case of a model failure”, is “enabled by default”, and kicks in when a model fails “due to quota or server errors”. It asks you first. The Gemini 3 docs describe the capacity case directly: when Pro “is overloaded”, the CLI asks whether to keep trying or fall back. “The Keep trying option uses exponential backoff… If the retry doesn’t happen immediately, wait a few minutes” (verified Nov 2, 2026). You can also run /model and pick Flash, or launch with --model. The /model docs say Pro is for complex tasks and Flash is for speed. On #4360, one user reported Flash working while Pro returned 503s.
  3. Check Google’s status page. Google AI Studio has a public status page at aistudio.google.com/status (responding Nov 2, 2026). If it shows an incident, step 1 is your only real move.
  4. Update Gemini CLI. Run npm install -g @google/gemini-cli@latest (troubleshooting docs, verified Nov 1, 2026). The maintainers closed #7227 asking people to reopen if it still happens on the latest version. The latest release was v0.66.0 on Oct 27, 2026 (verified Nov 1, 2026). This won’t add capacity, but it gets you the current retry handling and error text.
  5. Check your auth and project. Login with Google stopped working for individual accounts and Google AI Pro/Ultra on June 18, 2026 (codeassist.google, verified Oct 29, 2026), so individual users need another auth method, such as a Gemini API key. Rate limits apply per project, not per key (rate limits, verified Oct 30, 2026). /stats model shows usage per model. If it says you’ve hit limits, you’re in 429 territory, not 503.
  6. Report it, with details. If it keeps happening on the latest build, open or reopen an issue with your CLI version and the full error, as the maintainers asked.

Skip the “make a dozen projects and rotate keys” advice you’ll find in forums. That’s a quota workaround, and it doesn’t add server capacity anyway.

Does paying fix it?

Not reliably, based on what’s published. On #4360, users reported 503s on paid keys, including Tier 1 and Tier 3 keys (verified Nov 2, 2026). Google’s error docs treat 503 as server capacity, not as something tied to your tier.

There is a paid lever on the API side. Google’s Priority inference (Preview) says priority traffic “is prioritized above standard API and Flex tier traffic”. It costs “75-100% more than Standard”, and requests are set with service_tier: "priority". Overflow gets “downgraded to the Standard tier for processing instead of failing” (verified Nov 2, 2026). But Gemini CLI’s docs don’t expose a service-tier setting as of Nov 2, 2026, so whether CLI users can use it is not published. And a Google AI Pro or Ultra subscription no longer covers Gemini CLI since June 18, 2026. Paying for that to fix a 503 would be like paying for the gym to fix your car.

For what free and paid tiers actually get, see Gemini CLI free tier limits. If you’re shopping for something that doesn’t have Google’s capacity mood swings, Gemini CLI vs Claude Code and AI coding tools compared are the next stops. Every vendor has busy days, though.

Here’s what I’d do

Wait a minute and retry. If the CLI offers a fallback, take it. If Pro keeps failing, /model to Flash for the next task. Check the status page and update the CLI. If it’s still happening hours later on the latest version, file it with the full error. Don’t upgrade a plan to fix a 503 unless Google publishes something showing it helps.

If the error is a 403 rather than a 503, the Gemini CLI 403 permission denied fix covers Google login, API key and Vertex setups.

Getting 500 INTERNAL instead of 503 overloaded? See Gemini CLI Error 500.

FAQ

What does “The model is overloaded. Please try again later.” mean? Google’s servers lacked capacity for that model at that moment; it’s a 503 UNAVAILABLE server error, and Google says to wait and retry with backoff.

Is Gemini CLI error 503 my quota? No; quota and rate limits are 429 errors, while 503 means the service is “temporarily overloaded or down” per Google’s API errors page.

Does paying fix Gemini CLI 503? Not reliably; users on paid Tier 1 and Tier 3 keys report it too, and Gemini CLI doesn’t publish a priority-tier setting.

How long should I wait after a 503? Google recommends exponential backoff starting around 1 second; the CLI’s own “Keep trying” docs say to wait a few minutes if retries don’t go through.

Should I switch to Flash when Pro returns 503? It’s a reasonable move: Gemini CLI offers a fallback by default, and /model lets you pick Flash for speed.

Sources

  1. gemini-cli issue #4360, “The model is overloaded. Please try again later.” (full error text, paid-key reports, consolidated Sep 16, 2025): https://github.com/google-gemini/gemini-cli/issues/4360 (verified 2026-11-02)
  2. gemini-cli issue #7227, “Increased number of 503 errors” (collaborator note Jan 22, 2026, triage note May 9, 2026, closed May 13, 2026): https://github.com/google-gemini/gemini-cli/issues/7227 (verified 2026-11-02)
  3. gemini-cli issue #12932, “503. The model is overloaded.” (closed by stale bot Feb 15, 2026): https://github.com/google-gemini/gemini-cli/issues/12932 (verified 2026-11-02)
  4. Gemini API, API errors (503, 500, 504, 429, 400 definitions and actions): https://ai.google.dev/gemini-api/docs/api-errors (verified 2026-11-02)
  5. Gemini API, Troubleshooting (exponential backoff, jitter, which codes to retry, SDK retries): https://ai.google.dev/gemini-api/docs/troubleshooting (verified 2026-11-02)
  6. Gemini API, Priority inference (Preview; pricing, downgrade behavior): https://ai.google.dev/gemini-api/docs/priority-inference (verified 2026-11-02)
  7. Gemini CLI, Model routing (fallback on server errors, on by default): https://geminicli.com/docs/cli/model-routing/ (verified 2026-11-02)
  8. Gemini CLI, Model selection (/model, --model, Pro vs Flash): https://geminicli.com/docs/cli/model/ (verified 2026-11-02)
  9. Gemini CLI, Gemini 3 (capacity errors, “Keep trying” with backoff): https://geminicli.com/docs/get-started/gemini-3/ (verified 2026-11-02)
  10. Google AI Studio status page: https://aistudio.google.com/status (verified 2026-11-02)
  11. Gemini CLI troubleshooting (update command): https://geminicli.com/docs/resources/troubleshooting/ (verified 2026-11-01)
  12. Gemini CLI releases (v0.66.0, Oct 27, 2026): https://github.com/google-gemini/gemini-cli/releases (verified 2026-11-01)
  13. Gemini Code Assist consumer accounts (Login with Google ended June 18, 2026): https://codeassist.google/ (verified 2026-10-29)
  14. Gemini API, Rate limits (per project, tiers): https://ai.google.dev/gemini-api/docs/rate-limits (verified 2026-10-30)

Related reading