Gemini API key incident exposes a fragile OpenAI-compatible path

Google AI Studio’s live status page said on the June 29 check that Gemini API was having issues serving recently created keys, including when developers used those keys through OpenAI libraries. Google said investigation was underway.

The issue matters because Google’s Gemini docs explicitly support an OpenAI-compatible route: developers can point OpenAI Python or JavaScript clients at a Gemini base URL and use a Gemini API key. That is convenient for migration, testing, and multi-model routing. It also means a key provisioning problem can look like a model outage or adapter bug inside an agent stack.

What changed

• Google AI Studio’s status page flagged trouble serving recently created Gemini API keys.
• The status text specifically included use through OpenAI libraries.
• Google says OpenAI-compatible Gemini access works by changing the base URL and using a Gemini API key.
• Google’s API-key docs separately describe a 2026 shift toward auth keys and stronger key restrictions.

Buyer value

This is a reliability story for builders, not just a temporary status blip. Teams using Gemini behind OpenAI-compatible wrappers should know whether their production path depends on newly issued keys, old standard keys, restricted auth keys, direct Gemini SDKs, or a provider gateway. That makes it relevant to both AI Infrastructure and AI Coding teams that route model calls through shared SDK layers.

When a new key fails, the useful question is not only “is Gemini down?” It is also “which adapter path failed?” A direct Gemini SDK call, the OpenAI-compatible endpoint, and a third-party model router can fail differently.

What to do

For Gemini production use:

• run a smoke test every time a new API key is created;
• test direct Gemini SDK calls and OpenAI-compatible calls separately;
• keep older known-good keys out of rotation until replacement keys pass;
• tag alerts with the client path, model ID, region, and key type;
• avoid silent fallback to a different model unless the user-facing behavior is acceptable;
• document whether an agent framework uses Google’s SDK, an OpenAI-compatible adapter, or a ChatGPT-style client shim.

AiPedia take

OpenAI-compatible endpoints reduce switching cost, but they do not erase provider-specific operations. Key lifecycle, auth behavior, rate limits, and incident visibility still belong in the reliability checklist.