Hello. The free model everyone was reverse-engineering last week turned out to be GLM-5.3 Flash from Z AI, open weights, priced at a tenth of the frontier. Same week, The Information reported Nvidia has agreed to buy Hugging Face for $12.9 billion, which would put the place those weights live under the chip company. And Reddit's share of ChatGPT citations fell from 3.83% to 0.52% in four days, the first hard number on how fragile AI-search traffic is.

I also ran the bench I've been putting off. The cheapest model won. Details below.

Ox Alpha was GLM-5.3 Flash, and the price is the story

What happened. Z AI confirmed on Aug 26 that Ox Alpha, the anonymous free model that topped OpenRouter's usage chart for a week, is GLM-5.3 Flash. Weights are published, natively multimodal (text, image, video in one architecture), 320B total parameters with 18B active per pass. Z AI says the free week ran entirely on Chinese-made chips.

The details.

What I’d do. Nothing this week. Three points on an index don't mean three points on your work. Put it in a bench against whatever you use now (guide below) and let the results decide. If you have a workload with images in it and no budget for a vision model, this is the first one I'd try.

Source →

AI employees for small business

SUDO is our AI agency and product division, and it runs the systems behind this newsletter. Its AI employees answer email, chase invoices, keep the books tidy, and log a receipt for every action they take. Plans start at $199 a month, and the first two weeks are on us.

Nvidia is buying the library

What happened. The Information reported on Aug 26 that Nvidia has agreed to buy Hugging Face for $12.9 billion. Neither company has confirmed it and TechCrunch notes there is no signed contract yet. The same day, Nvidia reported a $96.2 billion quarter, up 106% on the year. Data center revenue was $89.0 billion, 92% of the total, gross margin 75.0%.

The details.

What I’d do. If your architecture assumes open-weight models stay free to download and cheap to host, write down what changes if the free tier gets a meter. Then check whether you can run your cheap tier on a box you own. I can, on a Mac Studio, and this week was the first time that felt like insurance rather than a hobby.

Source →

Reddit's citation collapse, and what it means for your AEO plan

What happened. Between Aug 14 and Aug 17, Reddit's share of ChatGPT Search citations fell from 3.83% to an average of 0.52%, according to Promptwatch. OpenAI said nothing. Promptwatch calls its own data provisional, and the likeliest trigger is a change on Aug 8 in how ChatGPT fans out queries: site-specific searches jumped from 0.37% of fan-out queries to 16.8% in one day.

The details.

What I’d do. I've spent the last month on answer-engine work for my own site and two clients. I'm keeping it, and here's the rule I'm adding: AI citations are a bonus channel, never a forecast line. Build the entity page, structure the content, ship the FAQ, then measure it as found money. Anyone selling you an "AEO strategy" with a traffic projection attached is selling you a number they can't defend. Ask them what happened to Reddit.

Source →

The flash bench

Every Monday my sweep script diffs the OpenRouter catalog against the week before, flags anything new that's cheaper than my current cheap tier with 128K-plus context, and writes a brief. Last week it flagged 19 candidates. I benched four: DeepSeek V4 Flash, Qwen3.8 Flash, Gemini 3.7 Flash, GLM-5.3 Flash.

The bench is six tasks shaped like my actual work. Classify a request. Pull fields from a messy email. Tidy a document. Summarize a long thread. Draft a client note. Judge whether another model's answer was any good.

DeepSeek V4 Flash won. It's now my cheap tier, my classifier (the thing that decides which model gets each job), and my judge. Qwen3.8 Flash and Gemini 3.7 Flash were close enough that I'd take either as a fallback and neither as a replacement. GLM-5.3 Flash I'll bench properly this week now that it's unmasked.

The part that matters if you own a budget: about 70% of what a business asks a model to do is on that list of six. The cheap tier now does it at a quality I couldn't tell apart from frontier in a blind read. The expensive models earn their keep on the other 30%, reasoning through a stack decision or writing something that has to persuade. If your vendor routes everything to one frontier model and bills per seat, you're paying frontier prices for clerical work. Ask them which model handles which task and who decided. "The best one, always" is a pricing strategy, not an architecture.

Also shipped this week: a client SEO report that used to run on a $200-a-month subscription now runs on a self-hosted stack with a data feed that costs about nine dollars a month per client. Same rankings, same backlinks, same audit, a better PDF, ten minutes per site. The saving is nice. The point is that I run it more.

A six-task bench for your own use cases

You don't need my tooling. You need a spreadsheet and an hour.

Tip. The judging step is the one people skip. Do it blind or you'll grade the expensive model up because you know what it cost.

A routing table for your company. Every model call goes through one file that says which tier handles which kind of task, what it should cost, and when to escalate. I run one for myself. I haven't seen a company run one. The first one that does will know its AI bill to the task, which nobody in your budget meeting can say today.

Distribution you own vs distribution you rent. Reddit just learned the difference. Your email list, your site, your sales team are owned. Search, social, AI citations are rented. Write down which of your channels is which, and what happens on a Tuesday when the landlord changes the locks.

The one-sentence handbook. Anyone who has tried to give an AI agent standing instructions ends up with a page of rules. I've been testing the opposite: one sentence describing the job, and letting the agent ask when it's unsure. Fewer rules, more questions, better output so far. Small sample. Playing with it.

Nothing else this week. I don't list tools I haven't run.

Got it wrong. I assumed free models on OpenRouter were unusable for real work. A finance-tuned model from inclusionAI passed my extraction test cleanly. Free is a tier now. Not for anything with a customer's name in it, but a tier.

Checked it. Two newsletters I read this week gave GLM-5.3 Flash different benchmark scores (57 and 87.4) and one reported the Nvidia deal as closed. It isn't. Every number above traces to a primary source, and I'll keep doing that. If you catch one that doesn't, reply.