Everything that happened in AI last week

Last week was one of the busiest week ever in AI in terms of drops and launches. This is a summary of everything that happened.

Models & Releases

Anthropic releases Claude Sonnet 4.5, now topping SWE-bench and coding benchmarks.
Google makes Gemini 2.5 Flash Image (“nano banana”) generally available with ten aspect-ratio options.
OpenAI launches Sora 2, a physically accurate text-to-video model and Sora social app.
DeepSeek unveils V3.2-Exp, a sparse-attention model that halves API costs.
Z.ai’s GLM-4.6 expands its context window to 200 K tokens and leads the open-weight LMarena leaderboard.
Qwen 3 Omni AWQ 30 B model released for 4-bit inference, boosting low-resource deployment.

Microsoft says future data-center AI workloads will run on its own custom chips, cutting dependence on Nvidia.
Nvidia announces the DGX Spark system will ship in October 2025, targeting large-scale LLM training.
OpenAI partners with Samsung and SK Hynix to secure up to 900 k HBM chips per month for its Stargate super-computers.
CoreWeave lands a $14 billion AI-infrastructure contract with Meta, expanding US data-center capacity.
MIT’s Lincoln Laboratory unveils TX-GAIN, a 2 exaflop AI supercomputer—the most powerful on any U.S. campus.
Granite-4.0-Micro (3 B) runs on Qualcomm Snapdragon X2 Elite NPUs, enabling on-device inference.

Anthropic’s Claude is now chat-enabled inside Slack for team collaboration.
Claude Code open-source agent brings terminal-based, context-aware coding assistance.
Google rolls out Jules CLI and API, letting developers run an AI coding agent directly from the terminal.
Perplexity releases the free Comet AI browser with a persistent assistant for web-search-enhanced workflows.
Onyx provides an open-source chat UI with built-in RAG, web search and multi-agent support.

California enacts SB 53, the first AI transparency law mandating safety disclosures and whistle-blower protections.
Meta will use data from its AI user interactions to target ads on Facebook and Instagram.
OpenAI updates its usage policies, restricting certain content and adding safety routing.
Reports confirm OpenAI now routes all users—including Plus and Pro—to lower-compute “5-chat-safety” models.

OpenAI’s new Sora app lets users create, edit and share AI-generated videos, a TikTok-style platform for short clips.
Google rolls out Gemini for Home, upgrading Nest cameras, speakers and adding AI-rich notifications.
Amazon launches a new Echo lineup powered by Alexa+ with advanced generative-AI features.
Sony updates WF-1000XM5 earbuds and WH-1000XM6 headphones with audio-sharing and Gemini Live AI assistant.
Apple pivots toward smart glasses, planning AR wearables that could launch as early as 2027.

Australian health agencies pilot AI support bots for home-care and diagnostic support, reporting higher patient engagement and reduced admin load.

“Radiology’s Last Exam” benchmark shows GPT-5 achieving only 30 % accuracy versus 83 % for board-certified radiologists, underscoring current limits of LLMs in medical imaging.