The model is especially good at agentic workflows, and in fast reasoning, Google says.
On the benchmarks it performs stunningly well for the price, even beating Gemini 3 Pro at MMMU-Pro, which tests for multimodal understanding. It also ticks in at 33.7% in Humanity’s Last Exam, just a little below GPT-5.2.
Google has priced the model for efficiency, too, with a cost of $0.50 for 1 million tokens of input, and $3.00 for the same output. This is substantially lower than 3 Pro and is only beaten by Grok 4.1 Fast and Gemini 2.5 Flash.
The model is rolling out in all channels as of today, and will become the default in the Gemini app and on the web, as well as in AI Mode, where it will much improve reasoning.
Read more: Google: launch page, for developers, and in AI Mode. Writeups at The Verge, 9to5Google and Engadget.