Google previews Gemini 3.1 Flash-Lite as its fastest and most cost-efficient Gemini 3 model

Google is targeting the economics of always-on AI workloads

On March 3, 2026, Google introduced Gemini 3.1 Flash-Lite and described it as the fastest and most cost-efficient model in the Gemini 3 series. The announcement is less about chasing the absolute top end of model capability and more about making high-volume AI work cheaper and faster to run. Google says the model is rolling out in preview to developers through the Gemini API in Google AI Studio and to enterprise customers through Vertex AI, which puts it directly in the path of teams building production systems rather than one-off demos.

Price and latency are the core of the pitch

Google priced Gemini 3.1 Flash-Lite at $0.25/1M input tokens and $1.50/1M output tokens. According to the company, the model outperforms Gemini 2.5 Flash while also delivering a 2.5X faster Time to First Answer Token and a 45% increase in output speed. Those numbers matter most for translation, moderation, routing, and other request-heavy services where small efficiency gains accumulate into meaningful changes in infrastructure cost and response time. Google is clearly positioning Flash-Lite for that class of workload.

Google is also making a quality argument

The company says Gemini 3.1 Flash-Lite reaches an Elo score of 1432 on the Arena.ai Leaderboard and outperforms similarly positioned models on reasoning and multimodal benchmarks, including 86.9% on GPQA Diamond and 76.8% on MMMU Pro. Google also says the model surpasses some larger Gemini models from prior generations. Real production adoption will depend on how those claims hold up in specific applications, but the message is clear: Google does not want Flash-Lite viewed as a stripped-down fallback tier.

Thinking controls widen the model’s role

One of the more important details is that Gemini 3.1 Flash-Lite includes thinking levels in AI Studio and Vertex AI. Google says developers can tune how much the model “thinks” for a task, allowing the same model to serve low-cost, repetitive workloads and more complex jobs such as generating user interfaces, building dashboards, creating simulations, or following detailed instructions. Google says early-access users including Latitude, Cartwheel, and Whering are already using the model. Taken together, the launch shows Google trying to compete on cost, latency, and configurable reasoning at the same time rather than treating low-price AI as a separate product category.

Source: Google