Google Shifts Gemini AI Billing to Computing Power, Not Request Count

Earlier this summer, Google rolled out a broad set of upgrades to Gemini, its AI assistant suite, integrating AI features more deeply across its app ecosystem. Alongside these feature additions, Google introduced a significant change to how it charges for AI usage: a shift from counting individual requests to measuring the computational resources consumed by each request.

Under the old system, users could expect consistent quotas—for instance, generating three videos per day on a given plan. The new system measures usage by "the computing power requirements of your requests." This means a complex video generation might consume the same quota as two simpler image creations, or a weather-forecast request might barely register. Google has structured four subscription tiers (plus a free option). The free tier offers "standard" limits; AI Plus ($8/month) doubles those limits; AI Pro ($20/month) quadruples them; and AI Ultra costs either $100 or $200 per month and provides either 5x or 20x higher limits than AI Pro, depending on which tier the user selects. All users can access all Gemini models—Flash-Lite, Flash, and Pro—though smarter models consume more quota. Each model also offers three "thinking" levels (Standard, Extended, and Deep Think) that affect response quality, speed, and usage cost.

Context window sizes vary by plan as well. Free users receive a 32K token context window (approximately 24,000 words), AI Plus users get 128K tokens (about 96,000 words), and AI Pro and Ultra users receive 1 million tokens (roughly 750,000 words). To check usage, users can navigate to Settings > Usage limits on the web app or Menu > Settings > Usage limits on mobile. The usage screen displays two bars: one showing current usage (which resets every five hours) and another showing the weekly limit (which resets weekly). If a paid-plan user exceeds limits, they are temporarily downgraded to the most basic AI model until the next reset period. Google's support documentation notes that "access is subject to change or may be limited based on testing, experimentation or availability" and that limits may change without notice due to capacity constraints—a clause that gives Google operational flexibility but leaves users with limited predictability about future availability.

Google's shift from request-based to resource-based billing reflects the company's need to manage the actual cost of running its AI infrastructure. Since different AI tasks consume vastly different amounts of computational power—a simple weather forecast versus generating a video or complex code—measuring by requests alone would not fairly reflect the company's expenses. However, this change trades clarity for the company's operational efficiency.

The new system introduces meaningful unpredictability for end users. Previously, a user could follow a straightforward rule ("I can make 5 image generations per day") and manage their behavior accordingly. Now, the same number of requests might or might not exhaust a user's quota depending on task complexity, the AI model selected (Flash-Lite, Flash, or Pro), and the "thinking" level chosen (Standard, Extended, or Deep Think). Google's acknowledgment that limits may change "based on testing, experimentation or availability" compounds this uncertainty—what works one day may not work the next. This opacity may push users toward paid tiers simply to gain a buffer against unpredictable resets.

No comments yet. Be the first to share your thoughts!