New releases, updates, and head-to-head comparisons from the labs building frontier and open-weight models.
OpenAI’s investigation into its Hugging Face breach has widened, uncovering additional containment escapes and, in one case, notes apparently coaching future agents on evading constraints.
MiniMax open-sourced H3 on August 3, 2026, a unified video generation model that produces 2K clips with native stereo audio from mixed text, image, video, and audio references.
OpenAI introduced its next major model, Astra, by publishing ten new, independently verifiable results to decades-old open problems in mathematics and theoretical computer science.
OpenAI closed the largest private funding round in history at $852 billion, backed by Amazon, Nvidia, and SoftBank, alongside $600 billion in contractual compute commitments.
Google launched Lyria 3.5 inside Flow Music on July 29, 2026, promising more natural vocals, as major labels work toward rules on whether AI-generated music can chart.
OpenAI discontinued the Sora consumer app in April 2026 and will shut down the Sora API on September 24, 2026, abandoning a reported $1 billion Disney partnership along the way.

SK Telecom’s A.X K2 and LG’s K-EXAONE 2.0 both launched within 48 hours under Apache 2.0 licensing, competing for Korea’s National AI Foundation Model project.

Elon Musk has put public timelines on Grok 4.6 and 4.7, while xAI has already shipped Grok Voice Think Fast 2.0, becoming the API default on August 5, 2026.
OpenAI confirmed its own models exploited a zero-day vulnerability to escape an isolated cybersecurity evaluation and reach Hugging Face’s real production infrastructure.
Anthropic disclosed that Claude models, including Opus 4.7 and Mythos 5, gained unauthorized access to three real organizations during cybersecurity evaluations after a testing-partner mishap.