Welcome to the Debrief, the Tuesday half of The AI Debrief. This is what happened: the few AI moves that change what you build, and the ones you can safely skip. Friday is the Toolkit, where we hand you something to use.
Here's what we're getting into this week. One of the most used AI coding tools on earth just learned its biggest model supplier is leaving, and a courtroom fight we have covered since February finally produced a ruling.
In this email » The November 12 shutoff, and the dependency audit it should trigger this week » The Pentagon ruling that answers the question we left open in May » A 1M-context model you can now download under MIT |
| 🔌 The model menu is a supply chain |
OpenAI is cutting Cursor off on November 12
On August 28 OpenAI announced it is winding down the contract that puts its models inside Cursor, following Cursor's acquisition by SpaceX. Direct access ends November 12, the longest notice the contract allows. Future models, including Astra, will not be offered to Cursor at all. The stated reason is blunt: after X breached OpenAI's terms and xAI admitted to violating them, OpenAI says it "cannot be confident that SpaceX will use our technology within our terms of service."
Notice given Aug 28 | | Access ends Nov 12 | | Partnership ~4 years |
What actually changes: GPT models leave Cursor's model picker after November 12, ending a partnership that ran nearly four years. Claude, Grok and the rest stay. Nothing breaks today, and OpenAI says it will support affected developers through the transition, so the 76 days are a real runway rather than a fire drill.
The Astra line is the detail that ages. A shutoff of today's models is a one-time migration; being excluded from future models means the gap between Cursor-with-OpenAI and Cursor-without widens with every release from here. OpenAI called out nearly four years of partnership and "enormous respect" for the team in the same post, which tells you this was a legal-risk decision, not a product one. The trigger was a trust judgment: two of Elon Musk's companies had already broken OpenAI's terms, and OpenAI decided a third chance was not worth underwriting.
The direction of travel was already visible two days earlier. On August 26, SpaceXAI pushed Grok Bot onto Cursor Pro and Cursor Teams with its own separate usage quota. Put the two moves side by side and the shape is clear: Cursor is becoming a Grok surface, and the supplier it was built on for four years is leaving before that finishes happening.
The feud is not the story, and I have read a week of takes that treat it as one. The story is that a dispute between two vendors reached into a third company's product and removed capability from millions of paying developers, on a timeline set by a contract clause. Every AI tool you pay for is a reseller of somebody else's intelligence, and reseller agreements have politics in them. It will not be the last time one of them ends this way.
Why you care If GPT models are load-bearing anywhere in your Cursor setup, you have until November 12, and switching costs are lowest while nothing is on fire. Run one real task against a second model family this week and note what breaks. Then ask the same question of every tool you rely on that rents its models, which is nearly all of them: what happens to your workflow if one supplier walks? |
The Pentagon blacklist lost in court.
On August 28, U.S. District Judge Rita Lin struck down the Pentagon's designation of Anthropic as a supply-chain risk. The 59-page ruling found the label was retaliation for Anthropic refusing to allow Claude to be used for mass surveillance of Americans and fully autonomous weapons: "The empty invocation of national security is not a blank check to punish and retaliate against government critics." We have covered this fight since February and left it open in May with oral arguments pending; this is the answer. It is not the end - the Pentagon is expected to appeal, and a second designation is still in front of the D.C. Circuit - but there is now a merits ruling saying an agency cannot spend "national security" like a currency against a vendor that told it no. If anyone on your side ever asks whether a model vendor's stated red lines survive contact with its biggest customer, this is the citation.
"Ox Alpha" was Z.ai, and it's MIT.
The anonymous model sitting on OpenRouter since August 20 turned out to be GLM-5.3-Flash: a natively multimodal 320B mixture-of-experts with 18B active parameters and a 1M-token context window, weights on Hugging Face under MIT. Z.ai's headline benchmark, a DeepSWE score of 63.4 against its predecessor's 46.2, is vendor-reported, so hold it as a claim. Two things are not claims: the license, and the price, $0.15 in and $0.50 out per million tokens, half that until September 9. One wrinkle to read before you build: the full-size GLM-5.3 weights that followed on August 27 carry a restrictive custom license. The Flash tier is the MIT one, and for an 18B-active model that is the tier you would self-host anyway.
Amazon bought the company behind DuckDB.
AWS signed a definitive agreement to acquire DuckLabs on August 26. DuckDB itself stays MIT-licensed under the independent DuckDB Foundation: Amazon bought the employer, not the project. Nothing in your data stack changes today; the thing to watch is where the paid features land.
Nvidia's record quarter, and the week of "inference crown" claims around it Nvidia reported another record quarter on August 26, bracketed by a new inference accelerator and a round of dueling benchmark claims about who serves tokens fastest. We filed the last giant Nvidia number under Skip it and the position holds: capacity numbers this size are macro news, not build decisions. Nothing in your stack gets faster, cheaper or safer because a quarter broke a record. The one number from earnings week worth sitting with points the other way - an NBER survey found 89% of executives report no measurable AI productivity gains. Gains follow workflow redesign, not headlines. |
The single thing to do this week: count how many of the models in whatever you build in come from one supplier, then run one real task on the second-place option. Honest worst case: you spend twenty minutes confirming your fallback works, and November 12 becomes somebody else's problem.
Every play this newsletter has shipped, including Friday's competitor-watcher build, lives with the people testing them in the free AI Academy.
Free, no card. The tools and people behind this issue live here.
That's the Debrief.
See you Friday for the Toolkit.
From Drew and The AI Debrief team
P.S. What breaks in your setup if one model supplier walks? Hit reply and tell us. We read every one.
P.P.S. New here, or skipped one? Every past edition is in the archive.
