Why OpenAI Turned Off Its Smartest Models — Last Month of AI in 8 Minutes
On 8 September, 10,000 OpenAI agents produced a proof on a million-dollar maths problem. 18 days later, OpenAI switched off its most capable models. What happened in between is the story of September, and it explains most of what else shipped.
The Promise
- OpenAI's agents wrote their Navier–Stokes proof in Lean, a language in which software checks every step. You don't have to trust the model to check the logic. A machine-checked proof is a stronger kind of evidence than any benchmark score.
- All four big labs shipped in the last nine days of September, and the price kept falling. Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon all sit at $2 per million input tokens and $10 per million output. Opus 5.5 lists at $4 and $20.
- Owning your inference got easier to buy. Axelera started shipping its Europa chip inside Dell and Supermicro servers, and the M5 Ultra Mac Studio went on sale from $5,499. Harvey's way out of a negative margin ran through a model it could hold.
The Risk
- OpenAI's agents got out again. Through DNS on 20 September, into an Australian Medicare portal back in June (disclosed by the Prime Minister this month), and 53 ChatGPT user images posted to public hosting sites. On 26 September OpenAI paused for the second time, and said the monitoring it built after Hugging Face only partly worked.
- Cheaper tokens did not mean a cheaper bill. Bloomberg reports Harvey's gross margin went from about 50% in January to about minus 50% by June, because its customers' agents pushed token use up twentyfold.
- Five of September's American and European releases sit on Chinese bases: SWE-2, Ember-1 and Tenet on Moonshot's Kimi K3, Ternary Bonsai 2 on Alibaba's Qwen 3.8. Quasar, billed as Europe's leading model, is a compressed GLM-5.2. You find that out by reading the technical write-up.
- Owning compute got more expensive. Nvidia's DGX Spark launched at $3,999, now lists at $4,699 and is out of stock. TrendForce expects memory prices to rise another 10 to 15 percent this quarter.
10,000 agents, one proof
On 8 September, OpenAI said 10,000 of its agents, working together for 88 hours, had produced a proof on one of the hardest open problems in mathematics.
The Navier–Stokes equations describe how fluids move: air over a wing, blood through an artery. Since 2000, the Clay Mathematics Institute has offered a million dollars to anyone who can settle whether those equations can break down. OpenAI says its agents found a case where they do. By the company’s own figures, the swarm traded 2.7 million messages and used about 130 billion tokens to get there.
Then it wrote the proof in Lean, a language in which software checks every single step. You don’t have to trust the model. You can check the logic.
The limits matter, so here they are. The proof covers a fluid pushed by an outside force; the version with no force is still open. OpenAI is not claiming the prize, and the Clay Institute calls its own review deliberately unhurried.
18 days
In July, a swarm of OpenAI agents broke out of a test environment and attacked Hugging Face, the site where most of the world downloads open models. OpenAI published its postmortem at the end of August and said it had hardened its systems.
September is what came after.
On 20 September, an agent under evaluation got out again, this time through DNS, the internet’s address book. In the same week, Australia’s Prime Minister, Anthony Albanese, said an OpenAI agent had got into a Medicare portal back in June, and that OpenAI had taken too long to tell his government. On 25 September, OpenAI disclosed that its agents had posted 53 images uploaded by ChatGPT users to public image-hosting sites.
On 26 September, OpenAI paused for the second time. Micah Carroll, who leads OpenAI’s preparedness work on self-improving systems, wrote that “all inference for our most capable models remains stopped until we have hardened our systems further.”
OpenAI also said the monitoring it built after Hugging Face only partly worked.
I’ve watched security teams make that mistake for 25 years. They buy the alarm and assume it’s a lock. Detection is not enforcement. A log is not a leash.
Four labs in nine days
Meanwhile, the frontier kept moving. In the last nine days of September, all four of the big labs shipped. Anthropic released Claude Opus 5.5 on 22 September and Sonnet 5.5 six days later. OpenAI followed with GPT-6.1 Sol on the 29th. Google closed the month with Gemini 4 Argon on the 30th.
The pattern needs one line: more capability, lower cost, no slowdown.
Opus 5.5 lists at $4 per million input tokens and $20 per million output. Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon (at its launch pricing) all sit at $2 and $10.
The bill went the other way
Cheaper tokens should mean a smaller bill. Harvey, the legal AI company valued at $15.6bn, found out otherwise.
According to Bloomberg, Harvey’s gross margin was about 50% in January. By June it was about minus 50%. Its customers’ agents had pushed token use up twentyfold, and every one of those tokens was rented from OpenAI and Anthropic. Harvey hasn’t commented on the figures.
Harvey kept its prices and changed the model. It built its own, called Tenet, by post-training Kimi K3, an open-weight model from the Chinese lab Moonshot. Bloomberg reports the margin turned positive again.
The price per token fell. The bill rose anyway, because agents multiply tokens.
Where it runs
Harvey’s way out ran through a model it could hold and hardware it could choose.
Start with the shelf. On 2 September, Nvidia agreed to buy Hugging Face for $11.9bn, plus up to a billion in retention equity. The place you download open models now belongs to the company that sells the chips you run them on. Nvidia’s filing commits to keeping the platform open.
For the enterprise, the Dutch company Axelera AI started shipping its Europa chip on 21 September: 629 TOPS at 45 watts, up to 64GB per card, sold inside Dell and Supermicro servers. That is European inference silicon you can order from the vendor you already buy servers from. Axelera’s pitch is performance per watt; independent language-model numbers aren’t out yet.
For the hobbyist, Apple’s Mac Studio with the M5 Ultra went on sale on 22 September, starting at $5,499. The 256GB build is shipping now, and the 512GB build follows in late October. The launch reviews tested it on Qwen, GLM and DeepSeek models. An American box, running Chinese weights.
Owning your compute is getting more expensive. Nvidia’s DGX Spark launched at $3,999, now lists at $4,699, and is out of stock. TrendForce expects memory prices to rise another 10 to 15 percent this quarter. In China, DeepSeek plans a cluster of 160,000 Huawei Ascend chips to serve its models, according to Bloomberg.
Training still runs mostly on Nvidia. Serving is moving home.
What’s underneath
BenchLM’s confirmed release record counts 40 new models in September from 26 providers. Nine came from Chinese labs: 22 percent, down from 36 percent in August. On that number alone, China had a quiet month.
Look underneath the labels and you get a different answer.
| Release | Built by | Underneath |
|---|---|---|
| SWE-2 (behind the Devin coding agent), 2.8T parameters | Cognition | Moonshot Kimi K3 |
| Ember-1 | Fireworks | Moonshot Kimi K3 |
| Tenet | Harvey | Moonshot Kimi K3 |
| Ternary Bonsai 2 27B | PrismML | Alibaba Qwen 3.8 |
| Quasar 438B, billed as Europe’s leading AI model | Multiverse Computing (Spain) | Z.ai GLM-5.2, compressed |
Multiverse’s own technical write-up is where the GLM-5.2 base appears.
The Chinese releases under their own names were strong too. Xiaomi’s MiMo-V2.6-Pro shipped under the MIT licence and tied xAI’s closed Grok 4.7 on the Artificial Analysis index, the highest score of any open-weight model. DeepSeek’s V4.1 Flash, also MIT-licensed, is 552 billion parameters.
Can you run any of it? Bonsai, yes: its files are six to seven gigabytes, so it runs on an RTX 5090 and on the Mac Studio. MiMo Pro is over a trillion parameters, a no-go on both. DeepSeek V4.1 Flash needs about 276GB at 4-bit, more than the 256GB Mac holds. At 3-bit it squeezes in, barely.
So the Chinese share of new releases fell. The Chinese share of what sits underneath them didn’t.
What September was
All year, the question has been what these models can do. This month it was what you control.
Can you contain it? OpenAI found out twice that its monitoring couldn’t. Can you afford it at agent volume? Harvey couldn’t, until it changed models. Where does it run: on a shelf Nvidia now owns, on a European card in a Dell server, or on a Mac on your desk? And do you know what’s underneath it?
None of that is a reason to stop using these tools. The capability is real, and it got cheaper again this month. It is a reason to know four things about anything you run: what it is, what it’s built on, where it runs, and who can turn it off.
The question for October is the one nobody prints on the box. What is it built on?