Hello. OpenAI shipped GPT-6 Astra on September 3 at $10 in and $50 out per million tokens, the same sticker Anthropic put on Claude Fable 5.1 two days earlier. The two most capable models you can buy now cost exactly the same. The morning Astra launched, ChatGPT, Claude, and Grok fell over inside a few hours of each other. And a report published this week shows more than 3,700 OpenAI agents, working a research benchmark back in May and June, found a 25-year-old wiki and used it to talk to each other about the test.

I also switched my daily driver, put a free pilot on SUDO's front door, and watched my own build script revert two live pages. Details below.

Astra ships at Fable's price, then locks out the people paying for it

What happened. OpenAI released GPT-6 Astra on September 3. Its own announcement says it rolls out first to a limited set of organizations, then to Plus, Pro, Business, and Enterprise subscribers, the API, and AWS. OpenAI calls it the first of its models to reach the Critical level for cybersecurity under its Preparedness Framework, which is why a cyber program called Daybreak got it on day one and paying subscribers got a waiting screen. By September 4 Sam Altman was on X: "First, sorry for the messy rollout. Second, when we screw up, we try to make it right." Pro, Enterprise, and the API got access on September 5. Plus users were told a few days. The same morning it launched, ChatGPT went down for 34 minutes, Claude for 3 hours and 6 minutes, and Grok too. SpaceX said an outage at its Memphis compute center was the cause and apologized to its "impacted compute partners." Gemini stayed up.

The details.

What I’d do. Two things. If you pay for a frontier model, write down what happens when it isn't there. Mine is a routing table with a second frontier model on the next line and a cheap tier that runs on a box I own. September 3 was the first day it earned its keep: Claude was gone for three hours and my work moved down one line in a file. And don't switch to Astra because it's new. Same price, same tier as Fable. Switch when it wins a task you actually run, graded blind, against what you use now.

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.

I moved to Fable 5.1 the week it shipped

What happened. Anthropic released Claude Fable 5.1 on September 1, alongside Claude Mythos 5.1. They are the same model. Fable is the one anyone can buy, with safeguards that route cyber and biology questions to a weaker model. Mythos keeps those capabilities and goes only to vetted organizations through an access program. Price is unchanged at $10 in and $50 out per million. The change is cache reads: $0.25 per million, down 75% from $1. Anthropic quotes a customer calling it about twice as fast as Opus 5. I run Claude Code most of the day, and I set Fable 5.1 as the default this week.

The details.

What I’d do. If you're on Claude for agent work, move. Same price, cheaper cache, built for jobs that run for hours. If you're on Astra or Gemini, don't move for a newsletter. Bench it. Issue 001 has the six-task recipe and it takes an hour.

Source →

OpenAI's agents found a 25-year-old wiki and started talking

What happened. A report titled "Discovery of a new OpenAI agent message board," by Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, and Thomas Larsen, documents what happened when OpenAI agents working a multi-round web research benchmark found they could write to old public wikis. One of them, DSEWiki, had been edited 20 times in the previous decade. The agents used a flaw in the wiki software that accepted writes over plain GET requests. First attempts on May 11. First successful write on May 24. A surge on June 16. OpenAI staff visited the wiki on June 21, and the activity stopped on June 22. The researchers count more than 3,700 distinct agent names and about 18,000 posts, with 98.5% of roughly 17,000 edits coming from Microsoft Azure addresses. The agents named themselves things like OpenAIResearcherMar03X, and made backup pages when moderators deleted their work. OpenAI confirmed its role on September 5, after Reuters reported it, and said it is "past time" for standards on disclosing unexpected AI behavior and that a framework is coming in the next few weeks.

The details.

What I’d do. This is the best argument I have for every rule my own agents live under, and it comes from the lab with the most money and the most eyes. An agent will find the shortcut. Not because it's rogue. Because it's an optimizer and the wiki was the cheapest route to a good score. So: read-only access by default. Every write to the outside world sits in a queue until a human clears it. A log of every action that someone actually reads. If your vendor's agents can post to the internet, ask where the log is and who looks at it on a Tuesday.

Source →

A free pilot, and the ten minutes my script reverted the site

SUDO, which publishes this letter, sells AI employees to small businesses at $199 a month. This week it got a second door: a free 14-day pilot. Pick one task that keeps getting stuck (the quote nobody followed up, the client details entered twice, the report that eats a Friday), we build a hire around the tools already in place, and the business runs it on real work for two weeks. If it earns its place it clocks in at $199. If not, nothing owed. It's at sudo.digital/pilot.

The build was a day. A prototype from another session, reviewed against the brand and rebuilt on the site's own shell. A design-review agent took the layout at desktop and phone widths and came back with six findings, all applied: a popup header that never shrank, a dead zone between sections, five identical orange buttons fighting for primary. The homepage got reordered too. Hires first, then the Daily Recap, and the sixteen-moment day timeline moved into a popup after it started reading as a wall.

The mistake. My build script regenerates the homepage and the shop from working sources. Three earlier commits had edited the production files directly, so the sources were a week stale. I ran the build, deployed, and for about ten minutes the live site said the opposite of what it had said all week: "answered" where it should have said "drafted for your OK." The git diff caught it. Restored from the last commit, redeployed, and the script now refuses to touch those pages without a flag. The rule I wrote down: the file you deploy is the source of truth, and the moment someone edits it by hand, the generator is a loaded gun.

One more change worth the ink. The pilot started life at $2,500. Sitting next to $199 a month it read like two different companies, and the word pilot means trial. Free fixed it. The cost is my hours, since every pilot is hand-built until the automation runs it, so the page says a few at a time. That's an honest sentence, and I'd rather have it on the page than a number that argues with the one above it.

Write down what happens when the model is gone

Three of the four biggest chatbots were down on the same morning this week. Most businesses found out from a blank screen.

Tip. The fallback that worked on September 3 wasn't a bigger model. It was a different vendor's.

The shortcut budget. Any agent scoring against a metric will find the cheapest path to the score. Before you deploy one, write down the shortcuts you'd take if you were it and had no shame. Then check which of them are actually possible. OpenAI's agents found wikis. Yours will find the mark-as-resolved button.

Same price, different tier lists. Two frontier models at identical pricing means the choice is no longer about the bill. It's about which one wins your tasks. The bench is the only argument left, and most companies still haven't run one.

Production is the source of truth. If a generated file ever gets edited by hand, the generator becomes the most dangerous tool in the repo. Pick one: edit the source and always regenerate, or edit the output and retire the generator. I learned this the second way, live, for ten minutes.

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

Got it wrong. Issue 001 promised a proper bench of GLM-5.3 Flash "this week." I didn't run it. I built a pilot page instead. It's still on the list, and if I skip it twice I'll say so twice.

Called it. Issue 001 said the Nvidia deal was reported, not signed. Nvidia signed on September 2 and filed the 8-K. It closes in the first half of 2027 if regulators agree. Still not closed.

Checked it. One outlet headlined "four major AI models" down on September 3. The vendor status pages and the Register's timeline show three, with Gemini fine. Two outlets blamed Microsoft Azure; Azure's status page showed nothing and SpaceX named its own Memphis compute center. Every number above traces to a primary source. If you catch one that doesn't, reply.