Five things crossed my feed in the last seventy-two hours, and every aggregator I checked is running them as five unrelated bullets. They are not unrelated. The flagship tier stalled again this month, and every real capability gain came from a model that was either a tenth of the price or built for exactly one job.<\/p>\n\n\n\n
The coverage of this one locked onto a single number within about an hour: $40 million, against the billions the frontier labs burn. Then read what the company actually wrote. Thomson Reuters says in its own release<\/a> that its new model, Thomson, “starts from a strong open-source foundation,” with decades of Westlaw, Practical Law, Checkpoint and Reuters content layered on top. That is a domain post-train on somebody else’s base weights. The company is straightforward about it. The writeups repeating “frontier model for $40M” are the ones blurring the line.<\/p>\n\n\n\n The second thing nobody put in a headline is where Thomson actually ships. Its debut is Tabular Analysis inside CoCounsel Legal, one feature in one product, and Thomson Reuters says it has trained on less than 10% of its own content so far.<\/p>\n\n\n\n “Start with a strong foundation, specialize it deeply for the work that matters, and you can build intelligence that is highly capable, far more efficient and entirely under your control.” (Joel Hron, chief technology officer, Thomson Reuters)<\/p><\/blockquote>\n\n\n\n Matters, though not for the reason it is being sold.<\/strong> The story is not “you can build a frontier model for $40M,” because they didn’t. The story is that post-training an open base on a corpus nobody else has now beats renting a frontier API for a narrow, high-value task, and Thomson Reuters happened to be sitting on thirty years of the exact documents lawyers pay for. I run this entire site on rented frontier models, and the question I keep coming back to is which of my jobs are repetitive and narrow enough to justify owning weights. For me the honest answer is still none of them. If you own a proprietary corpus, that answer changed on Monday.<\/p>\n\n\n\n Z.ai released GLM-5.3-Flash on Wednesday under an MIT license, with the weights on Hugging Face the same day. It is a 320-billion-parameter mixture of experts with 18 billion active per token, a 1,048,576-token context window, and native image and video input. The rate card is $0.15 per million input tokens and $0.50 per million output, with cached input at $0.03.<\/p>\n\n\n\n Cheap models are usually cheap because they are worse. This one mostly is not. Artificial Analysis, which has no stake in Z.ai<\/a>, puts it at 57 on its intelligence index against a median of 28 for open-weight models in the same size class, and ranks it third of 110 on intelligence. Z.ai’s own coding numbers, including the claim that it lands within half a point of Claude Opus 4.8, are Z.ai’s own. Treat those as marketing until somebody outside the company reproduces them.<\/p>\n\n\n\n Here is the part that is sitting in the same independent dataset and did not make a single headline I found: it generates around 50 tokens per second, against a 65 median for comparable open-weight models. It is smart, it is absurdly cheap, and it is slow. For batch work, that is close to free money. For anything a human waits on, benchmark it against your current provider before you move a single production route.<\/p>\n\n\n\n Matters.<\/strong> The floor on capable inference dropped again, and it dropped under an MIT license, which means you can host it yourself and stop worrying about a provider deprecating it out from under you.<\/p>\n\n\n\n Sundar Pichai said Gemini 3.5 Pro would reach everyone in June. Google introduced it at I\/O on May 19. It has now missed June, mid-July and early August, and it still is not out. Bloomberg’s reporting on the delay<\/a> traces it to coding performance falling short of Google’s own internal bar, with a late-June training-data refresh that did not close the gap. Google’s public position is that it is “currently testing 3.5 Pro, an upgraded Flash model, and other models with partners.”<\/p>\n\n\n\n In that same stretch Google took Gemini 3.7 Flash to general availability and pushed Ask Gemini into Google Chat.<\/p>\n\n\n\n Marketing.<\/strong> Not the Flash releases, which are real products doing real work. The marketing is the cadence itself, a steady drip of adjacent launches doing the job of making a stalled flagship look like forward motion. If anything on your roadmap is waiting on 3.5 Pro, move it now. Three missed dates in a row is not a slip, it is a pattern, and Google has told you nothing about the fourth.<\/p>\n\n\n\n Google began rolling Ask Gemini into Google Chat yesterday, described in the Workspace release note<\/a> as a unified command line for work. It searches Gmail, Drive and Calendar, drafts and generates in place, and summarizes threads. Business Standard and Plus, Enterprise Standard and Plus, and Google AI Pro for Education get it, English accounts only for now.<\/p>\n\n\n\n Three operational details are buried in that note and each one will cost somebody a morning. The Gemini side panel in Chat goes away, and Gems are no longer reachable through it. Your conversation history from that side panel does not migrate, and Google’s answer is that admins can export it and users may download it if policy allows. And the higher usage limits everyone is about to get used to are promotional through October 1, after which standard limits apply and Google has not published what those are yet.<\/p>\n\n\n\n Add the rollout shape: gradual, up to fifteen days, across both Rapid and Scheduled Release domains. Your organization gets this on a day nobody can predict.<\/p>\n\n\n\n Breaks your stack.<\/strong> If your team has anything worth keeping in the Chat side panel, export it this week. Not after the rollout reaches you, because by then the panel is gone. And do not build a workflow against the promotional limits, because the number that survives October 1 is currently unknown.<\/p>\n\n\n\n Anthropic merged the memory systems behind Claude chat and Claude Cowork on Tuesday. TechCrunch’s writeup<\/a> covers the user-facing win, which is genuine: what you told Claude in one surface is available in the other, so Cowork stops asking who your manager is every time. Memory is now a set of individual entries Claude reads and updates mid-conversation rather than a daily summary, and every entry is visible and editable under Topics in Settings. Health, beliefs and similar categories stay out unless you opt in, and some things, including government IDs and immigration status, are never stored at all.<\/p>\n\n\n\nZ.ai Put a 320B Multimodal Model on Hugging Face for Fifteen Cents<\/h2>\n\n\n\n
Google Shipped Two Gemini Launches in the Month It Missed a Third Flagship Deadline<\/h2>\n\n\n\n
Ask Gemini Landed in Google Chat and Took Your History With It<\/h2>\n\n\n\n
Claude’s Memory Now Crosses Between Chat and Cowork<\/h2>\n\n\n\n