Every development we logged over the past 30 days, what happened, and what it means for you. Newest first; each carries its cited evidence.
Apple's AI features can now launch in China, running on Alibaba's Qwen model.
DeepSeek's V4 line moves from preview to full release, with a very long memory.
Consumers lost access for hours, but OpenAI's business tools kept running.
Three models (Sol, Terra and Luna) now power ChatGPT and OpenAI's developer tools.
Artificial Analysis says Sol leads a coding test while using fewer resources.
OpenAI set Sol at about a third of a top rival's cost per task.
The tool tries to finish whole tasks (docs, decks and spreadsheets) on its own.
An outside lab placed Grok 4.5 among the top models, but not at the very top.
A small, open model lets robots follow spoken directions using a single camera.
New voice models aim to sound more human and handle back-and-forth talk better.
The company's newest AI is built for coding and multi-step tasks.
xAI set prices low to compete with pricier frontier models.
After a 19-day pause, the model returns with a filter that blocks a known misuse trick.
Meta wants to sell spare computing power and its models, like a cloud company.
The mid-tier model claims near-top quality at a lower price and powers Free and Pro plans.
Cheaper token prices run through August, though a change may raise how tokens are counted.
Regulatory clearance returns for Anthropic's most advanced models after a hold.
DeepSeek now charges more at peak times, a first for the fast-growing lab.
Nano Banana 2 Lite makes quick images; Gemini Omni Flash turns them into video.
Big-name departures to OpenAI and Anthropic raise questions about Google's AI lab.
Meta's first self-branded glasses put its Muse Spark AI on your face.