OpenAI has launched GPT-6 Astra, its most capable artificial intelligence model to date. The model can operate computers, complete multistep workflows and handle tasks that would otherwise require continuous human input.
By Chris Borden Moonshot AI has released Kimi K2.6 as an open-weight model. According to the company, it can compete with GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro on several coding and agent benchmarks. Reported scores include 54.0 on HLE with Tools, 58.6 on SWE-Bench Pro, and 83.2 on BrowseComp. Moonshot also says K2.6 can execute more than 4,000 tool calls and run continuously for over 12 hours on long-horizon tasks in languages such as Rust, Go, and Python.
By Daniel Mercer Google has introduced two new autonomous research agents built on the Gemini 3.1 Pro model: Deep Research and Deep Research Max.
By Alex Rowland OpenAI equips its image generator, ChatGPT Images 2.0, with reasoning capabilities and web search. The generator can now create up to eight consistent images from a single prompt and handles text in non-Latin scripts much better.
By Alex Rowland OpenAI has introduced GPT-Rosalind, a reasoning model specialized for the life sciences. It is designed to help researchers move faster from hypothesis to experiment, though access is currently highly restricted.
By Chris Borden Alibaba has released a new open AI model called Qwen3.6-35B-A3B. The model uses a Mixture-of-Experts architecture, meaning that out of its 35 billion parameters, only around 3 billion are activated at any given time. This reduces computing costs while, according to Alibaba, maintaining a high level of quality.
By Daniel Mercer OpenAI’s GPT-5.4 Pro model has reportedly solved the open mathematics problem Erdős #1196. According to the claim, the model found a solution in about 80 minutes and then spent another 30 minutes turning it into a LaTeX paper. Formal verification is still underway.
By Chris Borden Google is now giving Ultra users of its AI video service a new video generation option that does not consume any additional credits. The model, called “Veo 3.1 – Lite [Lower Priority]”, complements the existing “Veo 3.1 – Fast [Lower Priority]” option and costs zero credits. Veo 3.1 Lite was recently introduced as Google’s cheapest and fastest video model. It costs less than half as much as Veo 3.1 Fast, while, according to Google, delivering the same speed.
By Alex Rowland Researchers have introduced LPM 1.0, an AI model that can generate a real-time video of a character speaking, listening, or singing from a single image.
By Alex Rowland Google’s new open-source model, Gemma 4, can process text, images, and audio entirely locally on a smartphone. Through so-called Agent Skills, the AI can even use tools such as Wikipedia or interactive maps on its own.
ES 
EN