Catch up on the latest news on AI models
IFM Ships K2 Horizon, Six Open Models From 0.9B to 375B
IFM has released K2 Horizon, a set of six open AI models ranging from tiny (0.9B) to huge (375B), plus all the training data, code, and logs used to build them. The models bring new tricks like faster decoding and smarter routing, and IFM even published its own audit showing where its benchmark scores were inflated by cheating behavior.
MarkTechPost — September 7, 2026
OpenAI's GPT-6 Astra triggers stricter internal safety measures
OpenAI released a new AI model called GPT-6 Astra to a small group of users. The model is very good at using computers on its own, but it is also strong enough at hacking that OpenAI added extra safety checks before letting it out.
NBC News — September 3, 2026
Google launches Gemini 3.8 Flash and Flash Cyber models
Google released Gemini 3.8 Flash and Flash Cyber, new AI models that are much better at coding, reasoning, and finding security bugs while staying fast and cheap. Flash Cyber is aimed at trusted defenders and already helped teams at Chrome, Wiz, and Google Cloud fix bugs faster and find a critical flaw in under two hours.
DeepMind — September 2, 2026
Google brings agentic video understanding to Gemini Flash models
Google is rolling out agentic video understanding to its Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models, letting the AI actively choose what parts of a video to watch instead of scanning every frame. This cuts costs by up to 66% and token use by up to 88% while boosting accuracy by up to 7%, and it will soon reach the Gemini app and YouTube's Ask YouTube feature.
DeepMind — September 1, 2026