When an AI model is busy, retry twice with a growing pause, then switch to a backup only if you must. A 429 usually means you are sending too fast, a 503 that the provider is overloaded or down. Switching models also means starting without the prompt cache.
This is for founders and developers who run AI features in production, and for GrowApps users who see a "model is busy" message. It describes common practice and GrowApps' setup as of October 2026.
In short:
- A 429 is your limit, a 503 is the provider's capacity.
- Retries need pauses: back-to-back retries fail together.
- A new model usually reads your whole conversation uncached.
- GrowApps tells you before a backup switch that costs more.
Two errors mean two different problems
A rate limit and an outage look similar to the user but come from different places.
| Error | What it means | Usual fix |
|---|---|---|
| 429 Too Many Requests | You hit your quota or rate | Wait, retry, or use another model |
| 503 Service Unavailable | Provider is overloaded or down | Wait, retry, or use another provider |
| 500, or 529 on Claude | Temporary failure or overload | Retry with a pause |
A 429 often comes with a hint of how long to wait. Google's Gemini API, for example, documents its rate limits per model and tier.
Retries need a growing pause
The simplest fix is to retry, but timing matters. A demand spike can outlast several quick retries. Three retries in two seconds can all land inside it.
So space retries out and grow the pause each time, a pattern called exponential backoff. If the provider says how long to wait, honor it. Google Cloud's guide to retry strategy explains the pattern with jitter, a small random delay.
A same-family fallback suits rate limits
For a 429, a smaller model from the same provider often has spare quota. Switching to it inside the call can be invisible to the user.
The trade-off is quality. A smaller model can skip steps a larger one would take. In one session we reviewed, a smaller fallback model took no screenshots and shipped a layout overflow. Use this lane for short bursts, then return to the main model.
A different provider covers outages
When a provider is overloaded, its other models often share the problem. A backup from another provider is more robust.
It usually costs differently, and it starts cold. Prompt caches are per model, so the backup reads the whole conversation at full input price on its first call. On a long session that first call is expensive.
Different models also read prompts differently. In September 2026, a backup model in GrowApps read an automatic summary as an order to stop, and ended a build half done. We changed the summary's wording so the model carries on.
Users should hear about a costlier switch
A silent switch to a cheaper model is fine. A silent switch to a dearer one spends someone's money without asking. The honest pattern is to say what will happen, how much more it costs, and how to say no.
What GrowApps does when a model is busy
GrowApps handles both cases in production. On a rate limit, the agent retries the same model up to twice, waiting as the provider asks, up to 8 seconds. Then it falls over to a smaller model and cools the busy one down for about a minute.
On an overloaded model, the planner retries twice, after 3 and then 8 seconds. Then the chat says the backup costs more credits for this message, and that you can tap Stop.
After 15 seconds, that one message goes to a backup from another provider. Nothing is saved to your project, so your next message starts on the main model. Try a build on GrowApps.
Test your own retry path in ten minutes
Make one request fail on purpose, for example with a test stub that returns a 503, and watch what the user sees. Check that retries pause, that the message is plain, and that nothing is charged twice. Then check the bill for the turn, because a model switch also resets the prompt cache.
Frequently asked questions
Why does GrowApps not switch to the backup silently?
Because the backup costs more credits than the main model for the same message. We tell you before spending your credits on something you did not choose, and you can tap Stop to wait instead.
Is "the model is busy" a billing problem?
Usually not. In our experience it is most often the provider's high demand, a 503, which clears in minutes. Sending the message again shortly is usually enough.