According to monitoring by Dongcha Beating, Google plans to unveil its next-generation lightweight model, Gemini 3.2 Flash, at the I/O conference on May 20. The overall performance of the model is roughly on par with GPT-5.5, but it is clearly inferior to Anthropic's Mythos. Bindu Reddy, CEO of Abacus.AI, revealed rumors that Gemini 3.2 Flash achieves 92% of GPT-5.5's performance in coding and reasoning tasks, while its reasoning cost is only one-fifteenth to one-twentieth of the latter's, with most query latencies under 200 milliseconds. She believes that Google's distillation and sparsification techniques are playing a significant role, essentially compressing a cutting-edge model to Flash level without the usual performance cliff. There have been signs of leaks regarding Gemini 3.2 Flash. Earlier in May, traces of the model were found in iOS app build packages and AI Studio metadata, and it subsequently appeared anonymously in evaluations on LM Arena. Early testers reported that the model excelled in creative coding tasks, even surpassing Gemini 3.1 Pro in some benchmarks.
All Comments