Google’s latest open releases

Google spent the first week of April 2026 widening its “open” artificial intelligence lineup, adding a new Gemma 4 family, a cheaper Veo 3.1 Lite video model, and fresh Gemini API billing options. (blog.google) On April 2, Google said Gemma 4 is its newest open model family and released four sizes: Effective 2B, Effective 4B, 26B Mixture of Experts, and 31B Dense. Google said the models take text, image, and audio input, support context windows up to 256,000 tokens, and are available under the Apache 2.0 license. (ai.google.dev) (blog.google) Google said Gemma has now been downloaded more than 400 million times and has spawned more than 100,000 variants since the first release in February 2024. In the same April 2 post, the company said the 31B model ranked No. 3 and the 26B model No. 6 on Arena AI’s open text leaderboard as of April 1. (blog.google) A video model turns text or images into moving clips, and Google’s new Veo 3.1 Lite is aimed at developers who need more output for less money. Google Cloud introduced it on April 3 as the lowest-cost tier in the Veo 3.1 family, alongside the higher-end Veo 3.1 and faster Veo 3.1 Fast models. (cloud.google.com) In Google’s developer documentation, Veo 3.1 Lite Preview supports text and image prompts, returns video with audio, and drops some premium features, including 4K output and extension. Google said the model is meant for “high-volume video applications and workflows at scale.” (ai.google.dev) Google also updated how developers pay for Gemini API usage. Its pricing page now lists Standard, Batch, Flex, and Priority service options for models including Gemini 3.1 Flash-Lite Preview, while the billing page says new Prepay and Postpay billing plans began taking effect on March 23, 2026. (ai.google.dev 1) (ai.google.dev 2) The other release was TIPS, short for Text-Image Pretraining with Spatial awareness, a vision-language model family that tries to understand not just what is in an image but where things are. Google DeepMind’s paper says that matters for “dense” tasks such as semantic segmentation, where a model labels pixels or regions instead of giving one caption for the whole picture. (arxiv.org) Google DeepMind’s GitHub repository says TIPS combines contrastive image-text training with masked image modeling and uses synthetic descriptions to improve spatial understanding. The paper reports experiments across 8 tasks and 16 datasets, with the code and models released under an Apache-2.0 license. (github.com) (arxiv.org) Taken together, the April releases show Google pushing three lanes at once: open-weight models that can run on local hardware, lower-cost video generation for application builders, and research code that targets computer-vision tasks beyond chat. The next test is whether developers adopt Gemma 4 and TIPS as widely as they adopted earlier Gemma models over the past two years. (blog.google) (github.com)