Manus announced Wide Research on 31 July, an experimental mode that spins up more than 100 concurrent subagents to attack a single high-volume task, then reassembles their output. In the launch demo, co-founder and chief scientist Yichao Ji compared 100 sneakers, with each subagent assigned one shoe's design, pricing and availability, returning a sortable matrix as a spreadsheet and web page within minutes. It switches on automatically for tasks that need wide-scale analysis rather than long single-trajectory digging, and shipped first for Pro subscribers.

The architectural claim is the interesting part: Manus says this is the first application of a virtualization and agent layer that scales compute 100x beyond its earlier platform, with agent-to-agent communication handled at system level instead of through hand-built specialist subagents — the opposite trade from OpenAI, Google and xAI deep-research designs that spend minutes to hours on one trajectory.

Z.ai, formerly Zhipu AI, released GLM-4.5 on 28 July: a 355B-parameter mixture-of-experts model with 32B active parameters, alongside GLM-4.5-Air at 106B total and 12B active — base models, hybrid-reasoning models and FP8 builds all under the MIT license, downloadable from Hugging Face and ModelScope. Both are hybrid reasoning models with a thinking mode for complex reasoning and tool use and a non-thinking mode for fast replies, built as foundation models for agents rather than chat. On the company's twelve-benchmark average the flagship scores 63.2, which Z.ai places third overall and first among open-source models.

The price is the point: Z.ai lists API calls as low as $0.11 per million input tokens and $0.28 per million output tokens, with a high-speed tier above 100 tokens per second, while the weights themselves cost nothing to run on your own hardware. Z.ai reports 70.1% on TAU-Bench, 91.0% on AIME 24 and 64.2% on SWE-bench Verified with 128K context.

Microsoft launched Copilot Mode in Edge on 28 July, an experimental mode that puts one AI surface over the new-tab page and the page you are on, and — with your permission — reads context across all your open tabs. It is opt-in, free for a limited time for users in Copilot markets on Windows and Mac, supports natural voice navigation, and Microsoft says Copilot can take actions such as comparing options across sites and opening tabs for you. Microsoft says history and credentials will be handed over soon so Copilot can book reservations and run errands on the user's behalf, with visual cues whenever it is working in the background.

The differentiator in the AI-browser race is not chat, it is acting inside a session you are already authenticated to. That makes the browser the most privileged tool surface a consumer agent has yet been handed, and Microsoft is explicitly collecting consent per capability rather than per site.

Tesla signed a $16.5 billion contract to take its next-generation AI chips from Samsung, reported by Reuters and the WSJ on 28 July. Elon Musk had posted on X late on Sunday 27 July that Samsung's new Texas fab will be dedicated to making Tesla's AI6 chip, calling the strategic importance hard to overstate. Tesla's AI6 is its bet on a single all-in-one design that scales from supervised Full Self-Driving to the Optimus humanoid robot and up to data-center training.

The timing rhymes with the other 28 July headline: the same day 355B of agent-grade model weights went up for free under MIT, Tesla committed $16.5B to building accelerators. One half of the stack is being commoditized on price; the other is being pre-booked a generation ahead.