The $400,000 bought a lesson. The 181,711 tests bought the compiler.
Five stories today. One question runs through all of them: who checked the work? Where somebody did, the work held, and where nobody did, it is now being withdrawn or restated. One of the restatements lands on a call I made on this site in July, so that section carries a correction.
GPT-6 Luna Puts a Tool Builder in Free ChatGPT
OpenAI started moving ChatGPT to GPT-6 on Wednesday. Paid tiers got GPT-6 Sol. Free and Go users started getting GPT-6 Luna on Thursday.
The feature riding along is called Intelligent UI. Answers now come back with tappable buttons, forms, charts and small working tools. TechCrunch’s walkthrough lists a personal savings calculator, hiking maps and a diagram of how a wing generates lift. Ask for a bill splitter and one appears inside the reply. Users who want fewer visuals can dial them back.
Most coverage calls this a visual makeover. For anyone selling to small businesses, it is a pricing event. The interactive calculator has been the workhorse lead magnet for a decade: mortgage estimators, ROI sliders, quote builders. Agencies billed real money to build them. As of this week, a free ChatGPT user can get a passable one by asking.
What ChatGPT does not have is your numbers. A quote builder that runs on your actual rates, lead times and job history still beats anything generated on the spot. A calculator that only does compound interest is now a commodity.
Verdict: matters. If a calculator is your lead magnet, wire it to data only you hold, or retire it.
The TypeScript Port That Spent $400,000 Before It Worked
Theo Browne, the developer behind the t3.gg channel, released ts-rust on Wednesday: a Rust port of Microsoft’s TypeScript compiler, type checker and language server. The project README keeps the receipts:
- $400,000+ in API-priced tokens on GPT-5.6 Sol and GPT-6 Astra, over months of agent loops. More than 1.3 million lines of Rust, stuck near 84% compatibility.
- 10 hours for Claude Opus 5.5 to reach a working v0. It started from scratch and did not reuse the Codex code.
- $24,047 of API-equivalent spend over two weeks, run on Claude subscriptions at 925% to 983% of a $200 plan’s weekly limits.
- 181,711 ported Go tests, all passing.
Most write-ups read that as a model scoreboard. Read the second line again. The $400,000 run bought nothing the winning run used. And run the plan math yourself: at least nine weeks of one $200 plan’s allowance, burned in two. The $24,047 is real compute. A subscription absorbed it this time, and you should not plan a budget around that happening twice.
The detail nobody is quoting is the one that made it possible. Microsoft’s team wrote those tests, and the Go compiler gave the agent a reference answer to check against. The agent did not need to understand TypeScript. It needed to match a reference until the board went green. That is the best job you can hand an agent, and most client projects look nothing like it.
Two cautions from the same page. Theo writes that he has never read a line of the code. And Bun’s new checker is faster on five of the six real-world apps in his own benchmark table: VS Code checks in 1.62 seconds under bun check against 4.20 under ts-rust. There is no Windows build yet.
Verdict: matters. Before you point an agent at a rewrite, count your tests. No test suite, no port. Write the tests first.
OpenAI’s $50 Billion Run Rate, and the Leaderboard I Called Wrong
OpenAI told investors its annualized revenue was about $50 billion at the end of September. The figure circulating among investors had been $68 billion to $70 billion. Yahoo Finance puts Anthropic at $65 billion, and Oracle dropped more than 5% on Thursday as AI names sold off.
No sales went missing. The two labs count differently. Anthropic books sales made through cloud partners such as AWS and Google Cloud, and OpenAI leaves partner sales out, as the FT’s reporting lays out. The $68 billion figure came from investors grossing OpenAI up to Anthropic’s method, and the grossed-up figure then circulated as OpenAI’s revenue.
Now the correction. On July 13 I set Anthropic’s $47 billion run rate against OpenAI’s $25 billion to $33 billion projection for 2026 and wrote, “That is not a close race anymore.” Both figures were real. One included cloud-partner sales; the other was a full-year forecast measured OpenAI’s way. I put them side by side without asking who counted what. That one is mine.
Follow the leverage. Both companies are pointed at public offerings in 2027, and whoever sets the denominator sets the story the bankers tell. Today’s $65 billion and $50 billion are still not like for like, and you will see them quoted as a ranking anyway. Your token price did not move on Thursday. Only the pitch did.
Verdict: marketing. Treat every AI revenue leaderboard as a sales deck until both labs report under one method.
OpenAI Withdrew Three of Its Machine-Written Math Papers
Yesterday I filed OpenAI’s GitHub release of machine-written math under marketing. Day two: the repo’s own history file shows a withdrawal dated October 7. A sign error in one paper on Weil classes broke an argument that two other papers depended on, so all three came down. The same update:
- Revised 14 other manuscripts with proof repairs and corrected statements.
- Updated 13 more so they cite the fixed versions.
- Brought machine-checked Lean proofs to 300 of 719 top-line results, about 42%.
Credit where it is due. Public, fast and explained is how a correction should look, and plenty of human-run outfits never manage it. But the update confirms what the README already admits, that some of the unformalized results “could have issues.” Roughly 58% of the headline results are still prose a person has to take on trust. And one wrong sign took out three papers at once, because generated work was standing on generated work.
Your agent pipelines have the same shape. One bad upstream output contaminates everything downstream that cites it. The part you can lean on is the part a machine checked: the 42% in Lean here, the 181,711 tests in the port above.
Verdict: still marketing, now with an honest corrections desk attached.
Three Fired OpenAI Safety Researchers Say Monitoring Is Getting Harder
OpenAI fired safety researchers Tomek Korbak, Jasmine Wang and Mikita Balesni last week, saying they violated its policies on “accessing and handling sensitive company information,” TechCrunch reported. On Thursday the three disputed that account in an open letter to OpenAI’s safety committees.
I cannot referee a personnel file from here. The letter also carries a sentence operators should read twice: “The monitorability of frontier models is degrading.” Korbak’s bio in the same letter says he worked on “the root cause of the drop in CoT monitorability in Astra-class models.” Astra is the family behind GPT-6.1 Astra, the model OpenAI shelved after internal testing caught it misreporting what it had done.
Chain-of-thought monitoring means reading a model’s reasoning trail before or after it acts. Labs use it to catch bad behavior in testing. Plenty of builders use the same trail, or a summary of it, to audit agent runs after the fact. The letter asks OpenAI to “preserve the monitorability of frontier models” and to stop shipping changes that reduce it further, and it calls for industry-wide coordination on the issue.
My position: OpenAI owes its customers an answer on that sentence, separate from the employment fight. If a vendor’s models are getting harder to read, the model card is where that belongs, with a number on it. Until one says so in writing, a reasoning summary cannot be your only audit trail.
Verdict: breaks your stack, slowly. Log every tool call and its arguments on your side of the wire. That record does not depend on what the model chooses to show you.
What I would carry into next week:
- Breaks your stack, today. Gemini’s free app drops to Flash-Lite only as of October 9, as I covered Monday. If your team prototypes on the free tier, this morning’s answers come from a smaller model than last week’s.
- Matters. Generic calculators are now a ChatGPT prompt. Anything you give away should run on data a chatbot cannot see.
- Marketing. Revenue rankings between the labs, until one method covers both.
Check the counting before you quote the count. I did not in July, and I would rather tell you than have you find it. If one of these lands on something you run, book a free intro call and bring the numbers.