Key facts
- Google launched Gemini 3.6 Flash and 3.5 Flash-Lite, emphasizing efficiency and cost reduction.
- The Gemini 3.5 Pro model has been delayed due to internal performance issues, particularly in coding.
- Gemini 3.6 Flash offers a 17% reduction in output token usage and a lower API price compared to 3.5 Flash.
- Gemini 3.5 Flash-Lite is optimized for high-throughput applications and is significantly cheaper.
- Google is developing a custom AI chip, Frozen v2, for Gemini, with a target deployment around 2028.
- Pre-training for the next-generation Gemini 4 model has commenced.
Google has launched Gemini 3.6 Flash and 3.5 Flash-Lite, updated AI models designed for greater efficiency and lower costs, particularly beneficial for running AI agents at scale. Gemini 3.6 Flash shows improved performance on coding and engineering benchmarks compared to its predecessor, with a reduced output token count and API price. Gemini 3.5 Flash-Lite is optimized for high-throughput tasks, offering faster processing at a lower cost.
However, the release of the more powerful Gemini 3.5 Pro model has been postponed. It reportedly fell short of internal targets, especially in coding tasks, and attempts to rectify these issues have yielded disappointing results. This delay contrasts with the initial promise of a June delivery following its unveiling at Google I/O 2026.
In parallel, Google is reportedly developing its own custom AI chip, codenamed Frozen v2, specifically for powering Gemini models. This initiative aims for a 2028 deployment and is expected to deliver a six to ten times efficiency improvement over current Tensor Processing Units (TPUs). The development of custom silicon is seen as a move to reduce reliance on third-party hardware providers like Nvidia.
Google has also confirmed that pre-training for the next-generation Gemini 4 model has begun, signaling ongoing ambitious development in its AI capabilities. The company's parent, Alphabet, saw its stock price react to these developments, with a notable increase reported following news of the custom chip development.