Claude Gets a Browser of Its Own, and a Promised Weights Date Passes in Silence

This week was about what labs decided to keep inside. Anthropic gave Claude a browser of its own rather than reaching into the one you already have open. Z.ai let the August 28 date on its own Hugging Face page for GLM-5.3 pass without publishing anything or saying why, then released a smaller model openly instead. OpenAI finished narrowing its lineup by retiring o3. Only Google moved the other way, wiring one spoken request across Docs, Sheets, and Drive. Containment as a product decision, containment as a safety decision, and in one case containment with nobody obliged to explain it.

Tools and frameworks

Anthropic added a built-in browser to Claude Cowork on August 26, live on Enterprise and rolling out through the week to Pro, Max, and Team on macOS, Windows, and Linux. When a task needs the web, a browser opens in a side panel and Claude navigates, reads, clicks, and types inside it. The design point is whose browser it is: this one is Claude’s, isolated from your tabs, bookmarks, and logged-in sessions unless you deliberately move a login into it. Until now, giving Cowork web access meant handing it your own browser through the Claude in Chrome extension, which is still the better choice when the work is on a page you already have open. The distinction Anthropic is drawing is that most web tasks do not need your browser, just a browser.

Google went the opposite direction on the same day, connecting Gemini Live to Spark so a spoken request can run as a multi-step background job across Docs, Sheets, Drive, and the web. Spark holds the goal over days or weeks and keeps working while you are not in the app. Anthropic is narrowing the agent’s blast radius to a sandbox it controls; Google is widening it across every surface a user already lives in. Both are defensible, and they imply very different answers to what an agent should be trusted to touch.

Microsoft shipped an Excel Copilot update on August 25 that lets users reopen an earlier conversation rather than restart, and adds a change history skill that separates human edits from AI-generated ones. That second part is the useful bit, and it is the kind of provenance plumbing that gets boring and necessary at the same time.

Standards and open source

Z.ai’s Hugging Face repository for GLM-5.3 listed August 28 as the release date for the open weights. The repository sat as a placeholder for most of that day, and the weights landed late on August 28, which this briefing originally and wrongly reported as a date that had passed in silence. See the correction at the end. This is the model I wrote about last week: 84.5% on CyberGym, ahead of Mythos 5 and GPT-5.6 Sol, with exploit-chain reasoning Z.ai said its post-training produced without intent, held back for roughly two weeks of safety evaluation and hardening.

What Z.ai shipped instead, on August 26, was GLM-5.3-Flash, released under an MIT license and revealed to be the anonymous “Ox Alpha” endpoint that had been running on OpenRouter. It is 320 billion parameters with 18 billion active, and the native FP8 checkpoint runs about 306 GiB. So the open-weights commitment was honored, just not for the model the commitment was about.

Last week I framed the two-week hold as either genuinely load-bearing or a formality that produces a permanent outcome. On the evidence at the time it looked like neither, with the date apparently lapsing in silence. That reading was wrong, and the more interesting answer showed up in the license rather than the calendar. See the correction at the end.

Models

OpenAI retired o3 from ChatGPT on August 26, closing a 90-day sunset announced in late May. The API is unaffected. Because Canvas had already been pulled from GPT-5.5 Instant and GPT-5.5 Thinking earlier in the sunset, o3’s exit is the point at which Canvas effectively stops existing in ChatGPT, without a deprecation notice of its own. Retiring a model is a scheduled, announced event. Retiring a feature by outliving every model that carried it is not, and it lands on the same users either way.

Money and infrastructure

Stability AI raised a $76 million Series B on August 25, bringing its total to $232 million. The investor list is the story: Universal Music Group, Sony Music Group, Warner Music Group, and Electronic Arts, alongside AMD Ventures and Pacific Alliance Ventures. Three major labels funding a generative image and audio company reads less like a bet on returns and more like buying a seat at the table where training data terms get set.

What I am watching

Whether the terms attached to GLM-5.3’s weights get treated as openness or as something narrower. The pattern to watch is the one Z.ai is testing: ship the smaller sibling under a permissive license, and put the capable one out under terms of your own drafting. That satisfies the headline claim of openness at close to zero cost. If it works here without friction, it is a template, and “open weights” starts describing a license posture rather than an artifact anyone can use freely.

I am also watching the split between Anthropic’s isolated browser and Google’s cross-surface Spark. These are opposite bets on the same question, which is how much of a user’s live session an agent should be allowed to inhabit. Anthropic’s answer is easier to reason about and harder to make useful. Google’s is the reverse. The version that wins will set expectations for everyone shipping an agent that touches the web.

Sources

Correction

Updated September 5, 2026. This briefing reported that Z.ai’s August 28 date for the GLM-5.3 open weights “passed with no weights, no new date, and no statement,” and built its closing argument on that. It was wrong. The Hugging Face repository sat as a placeholder for most of August 28, which is the state the sources available on publication day described, but the weights were published later that same day. Z.ai met its own deadline.

The substance of the story turned out to sit in the license rather than the calendar. GLM-5.3 was released not under MIT, the license Z.ai used for GLM-5.2 and for GLM-5.3-Flash, but under a bespoke GLM-5.3 License that requires any model-as-a-service provider with more than $10 billion in revenue over any twelve month period to pass a security review by Z.ai. The 09-05 briefing covers what that means.