Google has released Gemini 3.1 Flash-Lite in preview, a model aimed at developers running high-volume workloads who want to control how much computational “thinking” each task receives. Positioned as the fastest and most cost-effective member of Google’s Gemini 3 line, it lets developers dial reasoning effort up or down rather than paying for deep reasoning on every request.
Adjustable thinking
The defining feature is a set of thinking levels — minimal, low, medium and high — that determine how much the model reasons before answering. For a simple, high-frequency task, a developer can choose minimal thinking to keep latency and cost low; for a harder problem, a higher level trades speed for depth. The control is available in both Google AI Studio, where developers build and test applications, and Vertex AI, Google’s platform for deploying machine-learning models at scale. Google suggests the model for high-volume translation, content moderation, and more involved jobs such as building user interfaces and dashboards or following multi-step instructions. It supports a context window of roughly one million tokens, with a maximum output of about 65,000 tokens.
Speed, price and strategy
Google lists the model at $0.25 per million input tokens and $1.50 per million output tokens, with reasoning (“thinking”) tokens billed at the output rate. On the independent Artificial Analysis benchmark it is reported to reach a first answer token about 2.5 times faster than Gemini 2.5 Flash, with roughly 45 percent higher output speed, while maintaining similar or better quality. Analysts read the release as characteristic of Google’s enterprise-oriented strategy: where some providers emphasise ever-larger reasoning models, Google is also competing on efficiency — cutting the tokens a model consumes to offer a capable but cheaper option. As one Futurum Group analyst put it, making a model markedly faster while nearly halving its price is a significant competitive move.
Mixing models
The launch also reflects a shift in how teams build with AI. Rather than relying on a single model, developers increasingly route tasks to whichever model fits: a more capable model such as Gemini 3.1 Pro for planning and complex construction, and a lighter one like Flash-Lite for routine documentation or straightforward code generation. The same logic of matching the model to the job appears across the industry’s recent model updates, including those covered in this look at OpenAI’s GPT-5.3 Instant.
What to watch
A few caveats are worth noting. Flash-Lite is in preview, so its behaviour, pricing and availability may change before general release. The headline speed figures come from a single benchmark and from Google’s own positioning, so real-world performance is best confirmed on the specific tasks a team cares about. And lower reasoning effort saves money precisely by thinking less, which can reduce answer quality on harder problems — making the choice of thinking level a deliberate trade-off rather than a free lunch. Full details are on the Google blog.
