Google's Gemini 3.1 Flash-Lite Is the Cheapest Capable Model Yet

Google released Gemini 3.1 Flash-Lite in preview this week, and the headline number is hard to argue with: $0.25 per million input tokens, $1.50 per million output. That’s the cheapest price Google has attached to a capable reasoning model, and the benchmarks suggest they haven’t cut the obvious corners.

On GPQA Diamond — the graduate-level science benchmark that punishes surface-level pattern matching — Flash-Lite scores 86.9%. It hits 76.8% on MMMU Pro, which tests multi-modal understanding across images and text. Both numbers beat larger Gemini models from the previous generation. The Elo score on the Arena.ai leaderboard sits at 1432, placing it meaningfully above models in its cost tier.

Speed is the other part of the story. Google cites a 2.5x improvement in Time to First Answer Token and a 45% increase in overall output speed compared to Gemini 2.5 Flash, based on Artificial Analysis benchmarks. For high-frequency workloads — content moderation pipelines, real-time translation at scale, bulk document processing — those numbers translate directly to infrastructure cost.

Adjustable thinking levels is the feature most developers will actually reach for first. Flash-Lite ships with configurable reasoning depth in both Google AI Studio and Vertex AI. You can dial down the model’s “thinking effort” for tasks that don’t need it, reducing latency and cost. Dial it up for more complex instructions. This kind of explicit cost-quality tradeoff is becoming table stakes for enterprise API models, and Google is building it in natively rather than as an afterthought.

The use cases Google lists are honest about the tier: high-volume translation, content moderation, UI generation, simulation workflows. This is not the model you reach for when accuracy on a single hard research question is paramount. It’s the model you reach for when you’re running a million calls a day and every 10ms and $0.001 adds up.

Flash-Lite sits at the budget end of the Gemini 3.1 family. Gemini 3.1 Pro, announced separately, targets the opposite end — deeper reasoning, higher context, higher price. Flash-Lite’s job is to keep Google competitive at the commodity layer, where Chinese labs like Alibaba (Qwen) and domestic API providers are pricing aggressively.

The model is available now in preview through the Gemini API in Google AI Studio and Vertex AI. General availability hasn’t been dated, but preview access is open.

The broader trend this represents is worth noting. Every major lab is now shipping distinct efficiency-focused models alongside their frontier work. The “one model does everything” era is over. What Flash-Lite signals is that Google intends to compete at every price tier simultaneously — not cede the low end while chasing performance on the high end.

At $0.25 per million tokens with reasoning controls built in, Flash-Lite gives teams a defensible reason to stay on Google’s infrastructure rather than defecting to cheaper alternatives.