Thursday's Rundown closed a weekend detective story: the mystery free AI that topped the charts was China's Z AI all along. I rewrote it the way I would tell a colleague over coffee. Short version: Ox Alpha is now called GLM-5.3-Flash, the weights are public, the price is tiny next to big rivals, and Z AI says a week of free use ran only on Chinese-made chips. OpenAI's boss put a date on "AGI". A small-business tip and a ChatGPT Work how-to round it out.
The newsletter also carried ads for Unwrap (customer feedback) and Memoket (a conversation wristband). Skipping those.
Words worth knowing
| Word | In one line |
|---|---|
| Model / weights | The trained AI. "Weights" are the saved numbers that make it work — publishing them means others can download and run it. |
| Open weights / open model | You can download the model and run it yourself, not only through the company's website. |
| Token | A chunk of text the model reads or writes. Not always a whole word. Pricing is often per token. |
| Frontier model | One of the current top-tier systems (the cutting edge of what labs sell or demo). |
| OpenRouter | A marketplace that routes requests to many AI models. "No.1 slot" means most used there. |
| Chip / silicon | The hardware that does the computing. Chinese-made chips = not Nvidia's usual cards. |
| Intelligence Index | Artificial Analysis's scoreboard for how "smart" a model looks on their tests (here: 57). |
| AGI | Artificial general intelligence — rough idea: AI that matches or beats humans on nearly every task. Nobody agrees on the exact bar. |
| Astra | A newer OpenAI model mentioned in the TIME piece; they claim it can do about a week of researcher work from a paper. |
| Agent | An AI helper that can take steps (click, run commands, draft files), not only chat back. |
| ChatGPT Work / Projects | ChatGPT's work mode that ties a chat to a folder on your computer so it can process files and set routines. |
| Skill / slash command | A saved routine you can run later by typing something like /name. |
| Post-mortem | A write-up after an incident: what went wrong and what they will change. |
The Ox Alpha mystery was Z AI's GLM-5.3-Flash
For about a week, developers chased an anonymous free model nicknamed Ox Alpha. Chinese lab Z AI has now said: that was their new model, GLM-5.3-Flash. They published the weights (the downloadable trained numbers) on Hugging Face and set a price far below comparable frontier rivals.
On OpenRouter (a site that routes traffic to many models), Ox Alpha took the No.1 usage slot. The newsletter says those numbers roughly doubled second-place DeepSeek, and OpenRouter ranked the debut as its biggest ever.
Z AI also said the free-usage week ran entirely on Chinese-made chips, and that this setup now serves tokens about as cheaply as Nvidia hardware. Independent lab Artificial Analysis grades the model at 57 on its Intelligence Index for a discounted $0.045 per task — about 10× cheaper than similarly ranked rivals, per the newsletter.
Why a GP should even care
Cheaper, downloadable models matter if a clinic tool's bill is mostly "how many tokens did we burn?". A strong open model on domestic chips is also a China story: less dependence on Nvidia. Treat the chip-cost claim as the company's, not a lab trial you can cite in a grant.
The Rundown notes Google staff had been publicly unsure whether it was Z AI — and that the original suspect turned out to be right.
Nate's Notebook: small business is AI's fast lane
Rundown educator Nate Grahek's weekly note: big companies move slowly; small outfits can still use that gap.
- Price side of the meter: Consumer/small plans from about $20–$200 are the subsidised ones. Enterprises often pay full freight per token. Stay on the cheap side when you can.
- Two lanes: People who already love AI should first buy back time (faster replies, less admin for your best clinicians), then mock up custom tools and hand them to someone technical to harden. People who never want to touch AI can still benefit if a small "tiger team" builds agents and automations that quietly clear boring tasks for them.
Next week he plans a part on individuals and how far AI can really take one person.
Practical: a beginner's guide to ChatGPT Work
ChatGPT Work can tie a project to a folder on your computer. Do the messy processing once, then turn it into scheduled automations.
- Make a local folder with a two-digit prefix. In ChatGPT desktop: New chat → Work → Choose project → New project. Name it and pick that folder as the source.
- Brief the project and ask ChatGPT to propose a structure. The guide's example: meeting notes → action items and weekly plans.
- Add input documents and tell it to process them once.
- Ask for three useful automations that run on a schedule so the same processing repeats without you babysitting it.
Pro tip
Ask ChatGPT to turn those routines into skills you can fire anytime with a slash command (like typing /weekly-plan). Same habit as a saved letter template: set it up once, call it when you need it.
OpenAI calls its AGI shot
In a TIME profile, Sam Altman put a date on AGI (artificial general intelligence — AI that matches or exceeds humans on nearly every task, depending who you ask). He said a model meeting his personal bar for that word will exist inside OpenAI by year-end.
- Research chief Mark Chen put the lab at "80% of the way" to AGI.
- Greg Brockman guessed people will look back on this stretch as AGI's arrival.
- Chief scientist Jakub Pachocki says their Astra model can take a paper and do about a week of researcher work alone — hitting an internal 2026 "automated intern" goal.
- Altman said Astra will be the first model that "actually invents new things in a way that matters."
TIME also covered July's security breach. OpenAI published its own deeper post-mortem and says it is refocusing on safety and slowing frontier development.
A claim, not a verdict
AGI is a flashy label with no shared definition. "By year-end inside the lab" is Altman's bar, not a regulator's, and not something that changes how you treat a patient tomorrow. The more concrete line in the piece is the automated-researcher milestone — still a company claim.
Quick hits
- Community workflow: Accountant Brian Carey used Perplexity's Computer agent to design a SharePoint document system for a biotech client — tagging rules, naming standards, 13 "START HERE" guides, then a script that built folders for 133 vendors (1,463 subfolders). Human still checked the counts.
- OpenAI called July's Hugging Face incident a loss-of-control "warning shot" and is keeping its biggest frontier training run on hold while it builds automatic shutdown for rogue agents.
- Bill Gates wrote that the world has "no plan" for the AI transition; he floated a tax on AI tokens and robots, and "Human Reserved" jobs set aside for people.
- ChatGPT Work can now sign in to websites via its browser for tasks — OpenAI says the agent never sees the user's passwords.
- Yutori released Navigator n2, a computer-use model that mixes clicking, terminal commands, and code.
- Anthropic opened Claude chat data to Stanford, Oxford, and METR for research; one group found over half of chats cover high-stakes topics like legal and financial advice.
- Trending tools named in the letter: GLM-5.3-Flash (the Ox Alpha reveal), Fal's H3 Max video model, Google Gemini 3.5 Transcribe (speech-to-text that cuts filler words), and Alibaba's Qwen3.8-Flash-Next preview.
What this means for a GP
None of this is clinical advice, and none of it should change how you treat a patient tomorrow. It is why a busy doctor might still skim the newsletter. A strong open model at a tenth of the price is how a practice-scribe or inbox tool stops feeling like a toy budget. Small-business advice that says "buy back time first, then harden the mock-ups" is the same discipline as a new EMR workflow: pilot on the boring tasks, do not hand the whole clinic to an untested agent. ChatGPT Work tied to a folder is useful only if someone still reads the action list before it goes on the record — same as a letter you did not type yourself. AGI deadlines make good headlines; the safer takeaway is OpenAI's own post-mortem language: when something can act on your behalf, you want an off switch before you want a miracle. Brian the accountant had the right habit: let the agent build the 1,463 folders, then check that the count matched.
Sources
- The Rundown AI, 27 August 2026 — AI's powerful mystery model revealed (Ox Alpha / GLM-5.3-Flash).
- Z AI blog — GLM-5.3-Flash / Ox Alpha confirmation.
- Hugging Face — zai-org/GLM-5.3-Flash weights.
- TIME, 26 August 2026 — Sam Altman / OpenAI AGI interview (Astra, 80% claim, year-end bar).
- OpenAI — Hugging Face incident post-mortem and the road ahead.
- The Rundown — Beginner's guide to ChatGPT Work.
- Bill Gates Notes — turbulent AI era / token and robot tax proposal.
- Anthropic — enabling independent research on Claude data.
- Yutori — introducing Navigator n2.