Google has released Gemini 3.1 Flash-Lite, which it describes as the fastest and most cost-efficient model in the Gemini 3 family. The model became available in preview on March 3, 2026, through Google AI Studio and Vertex AI, and it is aimed at developers who process very large volumes of requests each day.
Flash-Lite is the smallest tier in Google’s lineup, sitting below Gemini 3.1 Flash and Gemini 3.1 Pro. Google positions it for workloads where per-request cost and latency matter more than peak reasoning ability, such as classification, translation and content moderation at scale.
Pricing and speed
Google has set the price at $0.25 per million input tokens and $1.50 per million output tokens. That is well below the cost of the Pro model, but it is also higher than the previous Flash-Lite generation, a point noted by several outlets covering the launch. The Decoder calculated that the new model’s price is roughly triple that of its predecessor, so the cost advantage is relative to other Gemini 3 models rather than to Google’s older budget option.
On speed, Google says 3.1 Flash-Lite delivers its first token about 2.5 times faster than Gemini 2.5 Flash and generates output roughly 45 percent faster. Both figures come from Google’s own measurements.
Benchmark claims
Google reported a score of 86.9 percent on GPQA Diamond, a test of graduate-level science questions, and 76.8 percent on MMMU Pro, a multimodal reasoning benchmark. The company also cited an Elo rating of 1432 on the Arena.ai leaderboard, which ranks models by human preference in blind comparisons. These results place the model well above what budget-tier models achieved a year earlier, though they remain behind the larger Flash and Pro versions.
The model supports what Google calls thinking levels, a setting that lets developers control how much reasoning the model performs before answering. Lower levels keep responses fast and cheap for simple requests; higher levels trade latency for accuracy on harder ones. Google says this matters for high-frequency pipelines where most calls are routine but some need deeper analysis.
Intended uses
Google lists large-scale translation, content moderation, real-time user interface generation, dashboard and simulation building, and multi-step instruction following as the workloads the model was designed for. According to Google, early customers reported that it handled complex inputs with accuracy close to larger models at a fraction of the cost, though those are testimonials selected by the company rather than independent evaluations.
The release underlines a shift in how model makers compete. Alongside the race to build the most capable frontier model, Google, OpenAI and Anthropic are all now offering small, fast models priced for applications that make millions of calls a day, where efficiency rather than raw intelligence decides which provider a developer chooses.
