Nobody has a date for you
Microsoft's own AI went through Exchange Server and found so much that the update promised for this year now has no release date at all. Finding faults got cheap. Fixing them did not.
- 01
Microsoft cannot say when the Exchange update lands, and blames its own AI
In a post on the Exchange team blog last week, Microsoft explained why Cumulative Update 1 for Exchange Server Subscription Edition, first promised for the first half of 2026 and then for the second half, still has no release date. The reason it gives is its own AI tooling: the company has been running AI tools across its products to find vulnerabilities, and the Exchange team is now working through the resulting queue of issues to validate, reproduce, fix and test, which is eating the time CU1 needed. Microsoft's line is that CU1 is coming and it has not been forgotten, but it will only ship in a month with no pressing security work, and CU2 has moved into 2027.
Microsoft - 02
Three Imagen 4 models switch off today
Google's deprecation page puts 17 August 2026 on imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001, with gemini-3.1-flash-image named as the replacement. The notice went out in the changelog back in June, so anyone who read it has had two months. The point for a small firm is not the model: it is that if a supplier built you anything that generates images, today is the day it either kept working or stopped, and nobody sent you an email about it because you are not the API customer.
Google - 03
Stripe is reported to be buying the layer that lets you swap AI models
Bloomberg reported on 16 August that Stripe has agreed to buy OpenRouter for more than 7 billion dollars, roughly five times the valuation it raised at less than three months ago. OpenRouter is the plumbing a lot of small AI tools sit on: one account, one bill, and the ability to switch between models from different labs without rewriting anything. Stripe has not confirmed it and told reporters it does not comment on speculation, so read this as a deal reported rather than announced, but the thing being bought is exactly the thing that was supposed to keep you from being locked in.
Bloomberg - 04
Anthropic explains its text watermark, and what it does not prove
In a post on 14 August, Anthropic set out how the watermark works: the model's choice between equally good words is nudged by a private key, so no single word looks odd but the pattern shows up across enough text. It is a version of the SynthID-Text method Google DeepMind published in 2024, it adds nothing to your cost or speed, and it survives copy and paste and some editing. Two limits worth writing down. Only models launched from 2 August carry it so far, with older ones being retrofitted over the coming months, and the detection API is still promised rather than available, so nobody outside Anthropic can check anything yet.
Anthropic - 05
A 27B open model with vision, free to run on your own machine
Alibaba's Qwen team released Qwen3.8-27B on 14 August under the Apache 2.0 licence, weights and all, and it reads images and video as well as text. Quantised builds went up the same week in the formats people actually run at home and in small offices, and Simon Willison had it working on a laptop within two days, rating it excellent with the caveat that it overthinks simple questions unless you tell it not to. This is the one to watch if your objection to AI has always been that the drawings, the contracts and the client emails have to leave the building for it to work.
Qwen
Microsoft spent the last year telling everyone that AI would find the bugs. It has. Its own tools went through Exchange Server, turned up enough security work to fill the team's calendar, and the update promised for the first half of this year now has no date at all, with the one after it pushed into 2027. Nothing failed here. The finding worked exactly as sold. It is the second half of the job, validating and reproducing and fixing and testing each one, that has no AI in it and no more people on it than last year, and that half is now the whole bottleneck.
For a firm of twelve this lands somewhere dull and useful. Your suppliers are going to get slower at shipping the things they promised, while the monthly security patches keep arriving on time, because that is what the queue does to a roadmap. From direct experience in construction I know what a supplier saying "it's coming, no date" does to a programme: you stop believing the date, you build the week around what is actually on site, and you never let a promised delivery sit on the critical path. Same move here. If your plan for the autumn depends on a feature your vendor has announced but not shipped, take it out of the plan. And if someone is selling you AI code review on the basis that it will get your project out faster, ask who does the fixing, because Microsoft has just published what happens when the answer is the same team as before.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.