Seven Things And A Deadline

Loud week. Watermarks went global, the price of running AI moved more than it has in any two week stretch since these models showed up, an AI notetaker turned out to have been leaking meeting records since January, and a corporate breakup produced a deletion date that lands tomorrow morning.

Seven items, then the deadline. The deadline is the one with a clock on it, so if you're skimming, jump to the bottom first.

Anthropic began embedding invisible watermarks into text generated by its models on August 2. Not a visible label. A statistical fingerprint in the word choices, machine detectable, imperceptible to a reader. Files get signed provenance metadata using the C2PA standard instead.

The driver is Article 50 of the EU AI Act, whose transparency provisions became applicable that same day. Anthropic signed the bloc's Code of Practice on transparency of AI generated content alongside Google, Meta, Microsoft, OpenAI, Black Forest Labs and Synthesia. Non compliance carries fines up to fifteen million euros or three percent of global turnover.

What made this a story rather than a compliance footnote is that Anthropic applied the marking globally rather than just in Europe. No opt out. Every surface, including the consumer app, the API, Claude Code, and the versions running on AWS, Google Cloud and Microsoft Foundry.

What it means for you: less than the headlines suggest in the short run, more than you'd think over eighteen months. The mark records that a model processed the text. It says nothing about who did the thinking, and it degrades under paraphrasing, retyping, or passing text through a second model. Anthropic also hasn't shipped a public detection tool, so nobody outside the company can currently verify a mark on anything. Use the quiet period to write a one page position on where you use AI and what you disclose, before a client asks you cold in a meeting.

This is the item I'd have led with if the watermark story hadn't been bigger, because it hits your actual bill.

OpenAI and Anthropic have been cutting prices. DeepSeek went the other way and raised V4 Flash by ninety three percent, from fourteen cents per million tokens to twenty seven. GPT-5.6 Luna became the free ChatGPT default with unlimited chats after an eighty percent cut at the end of July. OpenAI also previewed Ultrafast, an API tier running GPT-5.6 Sol up to fourteen times faster on Cerebras hardware, at up to 750 output tokens per second, which is a bet that some buyers care more about latency than sticker price.

Underneath it, Anthropic reported second quarter revenue of 10.9 billion dollars, up 130 percent, and its first operating profit at 559 million. Companies cut prices from strength.

What it means for you: go read your last two AI invoices. Not the plan page, the invoices. Pricing in this category has become genuinely volatile in both directions, and the tool you picked nine months ago on price may no longer be the cheapest or the fastest option for what you're actually doing with it. This is a thirty minute review with a decent chance of finding real money, and almost nobody runs it because software costs feel fixed once they're set up.

xAI shipped Grok 4.6 on August 12 at two dollars per million input tokens and six per million output, the same headline rate as 4.5, with a 500,000 token context window.

Here's the part that wasn't in the launch post. Once a prompt reaches 200,000 tokens, xAI bills the entire request at four and twelve. Not the portion above the line. Every token in the request, starting from the first one. A request with 180,000 prompt tokens and 20,000 output runs about forty eight cents. Push the prompt to 200,000 and the same job costs roughly a dollar four. Cached input also rose from thirty cents to fifty per million, a sixty seven percent increase that didn't make the announcement either.

What it means for you: most of you aren't calling this API directly, so file it as a lesson rather than a task. The lesson is that headline pricing in AI increasingly describes a narrow band of usage, and the interesting money sits in the thresholds. If any tool you pay for bills by usage, go find its cliff. There's almost always one, and long running agent workflows are exactly the pattern that drifts across it without anyone noticing until the invoice lands.

Security researcher bobdahacker disclosed that tl;dv, a widely used AI notetaker for Zoom, Google Meet and Teams, left roughly 181,874 meeting records across 84,312 users and 35,003 email domains queryable by any authenticated user, because of a missing tenant isolation rule in its database configuration. About a thousand records were public, 715 invitee emails were exposed, and active recording sessions were joinable by outsiders, including calls involving government agencies and universities.

The researcher reported it in late January. It stayed unfixed through repeated follow ups.

What it means for you: this is the most actionable item on the page and it got a fraction of the attention the watermark story did, which tells you something about how tech news gets weighted. Go look at what your meeting recorder can reach and who on your team connected it to what. Notetakers sit inside your most sensitive conversations by design, and they usually get adopted by one enthusiastic person without anyone reviewing the security posture. That isn't an argument against the category. I use one. It's an argument for knowing what you turned on.

Two releases pointing the same direction. Meta put out Muse Glimmer, a thirty billion parameter multimodal model under Apache 2.0, tuned for local agent work and coding, with a 131K context window. Quantized to four bits it fits under twenty gigabytes, which puts it on a single consumer GPU. Alibaba's Qwen team released Qwen 3.8 27B, also Apache 2.0, with integrated vision and a 262K native context.

What it means for you: the use case here is narrower and more useful than the hype suggests. You're not replacing your main model. You are running the work that shouldn't leave your building. Client data covered by a confidentiality agreement, internal documents you'd rather not send to a third party API, anything where the vendor's terms make you hesitate. If you've been declining to use AI on your most sensitive work because of where the data would go, this is the year that stops being a good reason.

Mercury launched Mercury Spend, which includes budgets, self enforcing policies, and a card an AI agent can actually use to make purchases end to end, receipts included.

What it means for you: fascinating, and not yet. The direction is right, because an agent that can research a purchase but can't make it has produced a to do list. But an autonomous system with spending authority is a risk category most small businesses have no controls for, and the failure modes are new enough that nobody has good instincts about them. If you experiment, cap it low, scope it to one vendor category, and check it weekly. The people who get hurt here will be the ones who set it up correctly and then stopped looking at it.

Adobe, working with Morning Consult, surveyed 850 US small business owners. More than half said that when everyone has the same software, the advantage goes to whoever has the sharper eye. Seventy nine percent said a clear point of view is what makes people follow them. Nearly one in three would rather post nothing than sound off brand.

Sixty six percent already use AI for marketing. Only six percent published their last AI draft untouched, and forty eight percent reworked it heavily.

What it means for you: that six percent is the most useful number I've seen this year. Ninety four out of a hundred owners independently decided that raw model output wasn't good enough to carry their name, without anyone handing them a framework. They're right. The move isn't to use AI less. It's to use it on the parts that are identical for everyone, which is structure and formatting and cleanup, and keep the parts that only exist because it's your business, which is the numbers, the names, the mistakes and the opinions.

Manus told users on August 11 that it's returning to independent operation. Meta acquired it in December for a reported two billion or so. In April, China's National Development and Reform Commission blocked the deal and required the parties to unwind a transaction that had already closed, which almost never happens.

The cleanup includes deleting data generated by affected users on or after December 29, 2025, the acquisition date. Manus says this is regulatory compliance, not a security incident, and that anything created before the acquisition is untouched.

The backup window closes at 7:59 in the morning Singapore time on August 23. That is roughly this evening on the US East Coast. Deletion runs from 8:00 that morning through the 24th, and restore opens on the 25th. Manus is waiving charges during the backup window and notifying affected users in the app and by email, though anyone who signed up with an Apple ID or a Facebook account only gets the in app notice, which is the kind of detail that guarantees somebody misses this.

If you or anyone on your team has used Manus since December, go export now and read the rest of this later.

And if you don't use Manus, which is most of you, the lesson still applies. Nobody with work in those accounts made a bad vendor choice. They picked a startup that got acquired by Meta, which is about as establishment an outcome as a startup can have. Then two governments had an argument and a deletion date appeared on their calendar. You cannot research your way around that. You can only build the export pipeline before you need it, which is exactly what we built on Tuesday. If you skipped it, the short version is a monthly scheduled job in Make that pulls your records to dated folders and only messages you when the file comes back smaller than last month.

Two threads this week, and they're worth separating because they call for different responses.

The first is control. The watermark is about provenance. The Manus deletion is about custody. The tl;dv leak is about access. The local model releases are about where your data physically sits. Every one of those is a question about something you'd previously assumed was just handled by somebody competent.

The second is cost. Prices moved hard in both directions, a major launch buried a repricing cliff in a footnote, and a new tier appeared that sells speed instead of intelligence. The era where you picked a tool, learned its price, and stopped thinking about it is over for a while.

Neither thread requires you to slow down or stop adopting anything. Both require you to go back and look at what you already built. That's the unglamorous part of this phase, and it's where the businesses that come out of it well will separate from the ones that just kept adding tools.

Three answers, and you can have all of them by Wednesday. Where does your data live. Who can reach it. What are you actually paying.

One thing this weekend. Open your meeting recorder and look at what it's connected to and who connected it. Of everything on this page, that's the item most likely to be quietly true in your business right now.

Tomorrow, the Sunday piece, and it's about a list I think every operator should write and almost nobody has.

Jordan

The AI Newsroom is written by Jordan Hale. This issue contains affiliate links to tools I actually use. If you sign up through them I may earn a commission at no extra cost to you.