The Gemini family of models is getting a new member. This time, the company is introducing Gemini 2.5 Flash-Lite, in preview, starting today, alongside making the more powerful 2.5 Pro and 2.5 Flash models generally available for production use.
Google notes that Flash-Lite is better than its 2.0 predecessors across coding, math, and reasoning benchmarks. The main idea is for it to handle things like translation or classification at a massive scale, where speed is essential. The pricing certainly reflects this focus on volume. Flash-Lite is the cheapest option by a wide margin, costing $0.10 per 1M input tokens and $0.40 per 1M output tokens, but its baseline performance on difficult reasoning tasks is quite low, scoring just 5.1% on Humanity's Last Exam (full technical report here).
The most interesting feature, however, is its "Thinking" mode. When you turn this on, the model takes more time to process a prompt, and its capabilities improve dramatically. For example, its performance on a difficult mathematics benchmark called AIME 2025 jumps from a respectable 49.8% to a much stronger 63.1%. The same pattern appears when testing its ability to find information in a large document, where its recall ability nearly doubles. Here's the benchmark table Google provided (click to expand):
Google says developers can start experimenting with this new model immediately. The preview version of Gemini 2.5 Flash-Lite is now accessible through Google AI Studio and the developer-focused Vertex AI platform. The more powerful, stable versions of 2.5 Flash and Pro are available in those same places but are also integrated into the main Gemini app for consumers. The company even noted that custom versions of both Flash-Lite and Flash are already being used to power parts of its own Search engine.
1 Comment - Add comment