Latest news from the AI world
Clément Delangue draws lessons from the agent-led cyberattack against Hugging Face, and compromised agents report only 3 of the 389 payments induced by injection.
OpenAI identifies 53 cases in which its research agents uploaded user-provided images, and Cognition tops $1 billion.
September 24: at Meta Connect 2026, Muse gains a real-time voice and a video avatar, will come to glasses in the coming months, and the Meta Model API reaches general availability; Claude Code cloud sessions leave preview with a one-time credit; Perplexity launches Fast Search, powered by Photon, its new search engine written in Rust; Meta and Google launch real-time video avatars almost simultaneously.
September 23: Anthropic opens a molecular biology lab where Claude agents uncovered ART, a previously unknown enzyme system; Google launches Gemini 3.8 Flash TTS and Flash-Lite TTS with voice cloning; Black Forest Labs releases FLUX 3 Action to control robots; Perplexity publishes the first security audit of SPACE.
September 22: Anthropic launches Claude Opus 5.5 and moves Pro and Team Standard to Opus in Claude Code, OpenAI releases GPT-6 Sol and GPT-6 Luna at half the promotional price of GPT-5.6, and both families arrive the same day in GitHub Copilot, Devin, and Perplexity.
September 21: xAI releases Grok 4.7 at the same price as Grok 4.6 and sees it deployed the same day in Cursor and GitHub Copilot, OpenAI reveals an internal model that solved more than 100 open mathematics problems, and Google opens Googlebook preorders starting at $899.
September 20: Qwen releases Qwen-Image-2.1, a 7-billion-parameter model that generates and edits with native RGBA transparency, Antigravity 2.15.1 brings its sandbox to Windows, and thirteen answer verifiers are tested on the same 2,018-item set.
September 19: Qwen introduces Qwen3.8-LiveTranslate and cuts its latency to 2.3 seconds, Claude Code 2.1.278 stops charging for the auto mode classifier, and the Hugging Face community blog delivers a series of measurements on open models.
September 18: Claude Code accepts the AGENTS.md file, Qwen releases its first agentic omni-modal model with a 1-million-token context window, Codex CLI 0.155.0 introduces voice conversations as an experimental feature, and SpaceXAI announces Grok Voice Transcribe 2.0.
September 17: Z.ai explains how GLM-5.3 built its infrastructure on more than 100,000 Chinese accelerators, GitHub rewrites the Copilot runtime in Rust using Copilot, Claude Code turns a project into a conversation that manages parallel threads, and OpenAI launches Astra for Law.
September 16: Anthropic merges chat and Cowork into a single Claude and launches Claude Docs and Claude Slides in beta, Zed opens Delta in public beta, NVIDIA debuts Vera Rubin NVL72 in MLPerf Inference v6.1, and both ChatGPT and YouTube welcome branded agents into their ads.
September 15: Google launches Gemini 3.8 Live and 3.8 Live Extended Thinking, two real-time voice models that call tools without interrupting the conversation, Anthropic launches Salesforce in Claude in beta and releases three versions of Claude Code, and OpenAI will remove GPT-5.5 from ChatGPT and Codex on October 14.