Why Free AI APIs Are Becoming Harder to Rely On
On December 6, 2025, developers started waking up to “quota exceeded” errors. Apps that had worked for months suddenly stopped working because Gemini 2.5 Pro had disappeared from Google’s free API tier.
Eight months later, the outage itself is old news. The explanation Google gave for it is more useful.
Logan Kilpatrick, Google’s product lead for AI Studio and the Gemini API, said the generous free access had “originally only supposed to be available for a single weekend.” Instead, it lasted roughly seven months, long enough for developers to build projects around it. When Google needed capacity for newer models, that subsidy disappeared.
The problem wasn’t the price
Gemini 2.5 Pro’s paid API pricing was public: about $1.25 per million input tokens and $10 per million output tokens for shorter prompts at the time. Developers could have budgeted around those numbers.
What they couldn’t budget around was not knowing whether the free tier would still exist when their projects reached the next stage.
Kilpatrick was explicit about the purpose of the free tier. It was meant to give developers “a small taste of the model,” not serve as infrastructure for ongoing products. That’s a reasonable business strategy.
The problem is that seven months of generous availability can look very different from a temporary promotion to someone actually building on it.
Free AI was always a subsidy
Frontier AI is expensive to run. Every API request consumes computing capacity that providers can instead allocate to paying customers or newer models.
A generous free tier therefore serves a purpose beyond goodwill. It gets developers experimenting with a model, learning its behavior and building products around its ecosystem. Once that adoption has been established, the provider has less reason to keep absorbing the inference cost.
That doesn’t mean every free tier is about to disappear. It means developers should stop treating free access as a stable infrastructure layer.
Other providers are facing the same economics. DeepSeek, for example, introduced peak-hour surcharges for some models in 2026 after years of competing aggressively on price, citing resource allocation and service stability. The details differ, but the underlying pressure is the same: AI compute is not free, even when the API is.
What developers should change
For prototypes, free tiers remain useful. Students, hobbyists and developers testing whether a model fits a particular task can still get substantial value from them. Production is different.
The first calculation should be what the application costs at the provider’s published paid rate. If the economics only work because the API currently costs nothing, the economics don’t really work. The second is portability.
AI models are not interchangeable like database endpoints. Switching providers can mean retesting prompts, output formats, tool calls and model-specific behavior. But building an abstraction layer early can turn a provider migration from a rewrite into a configuration and testing exercise.
That’s increasingly important as providers adjust limits, pricing and model availability around changing compute demand.
The lesson from Google’s weekend
The developers affected by the Gemini 2.5 Pro change weren’t wrong to use the free tier. They were wrong only if they assumed its availability represented a commitment. That’s the distinction worth carrying forward.
A free AI API is best treated as an evaluation period with an uncertain expiry date, not as the foundation of a product’s cost structure. Google didn’t make frontier AI suddenly expensive in December 2025. It simply stopped hiding the cost behind a subsidy.
