Yesterday a California judge vacated the Pentagon’s Anthropic supply-chain label. Overnight the legal story sat still. The coding-tool one did not.

OpenAI notified SpaceX it will wind down the contract that supplies OpenAI models to Cursor, with a proposed shutoff of November 12, 2026. That date, the company says, is the maximum notice the custom agreement allows after a change of control. SpaceX closed its 60 billion all-stock purchase of Anysphere — Cursor’s parent — on August 14.

The reason is not “we do not like the product.” OpenAI says it cannot be confident SpaceX will stay inside the terms of service, citing Musk companies that already broke them. After Musk bought Twitter (now inside SpaceX), OpenAI says the company broke that contract. Under oath earlier this year, Musk admitted that xAI — also now part of SpaceX — distilled OpenAI data against similar terms. The upcoming model Astra is named in the same post: Cursor will not get it.

Cursor co-founder Michael Truell, now at SpaceXAI, put the damage at about 5 percent of user traffic and said the team is still talking to OpenAI. Anthropic went the other way. Co-founder Tom Brown said Cursor has been a partner since Sonnet 3.5 and that Anthropic will increase compute for Claude inside the editor. Musk’s reply, per Reuters: he could not care less.

If you actually live in Cursor, that 5 percent is the number that matters. The rest of the stack is already Claude, Gemini, and Grok. November 12 is when the bundled OpenAI path goes dark, not when coding stops. The interesting part is the clause, not the traffic: change of control is now a model-access event. Labs will write that into every IDE contract from here.

A used workstation and a second monitor still beat an IDE that can lose a frontier model because someone sold the company.

GLM-5.3 weights actually landed

Z.ai promised open weights about two weeks after the August 14 API launch. They shipped. GLM-5.3 is on Hugging Face: 744 billion total parameters, 40 billion active, same GLM-5.2 base, all of the lift from post-training. Context is 1 million tokens. Unsloth already has GGUFs; the UD-IQ2_M quant wants on the order of 256GB of unified memory or a lot of RAM plus modest VRAM. This is not a laptop model.

The license is not MIT. GLM-5.2 was. GLM-5.3 uses a custom GLM-5.3 License: if your trailing twelve-month revenue is over 10 billion, you need Z.ai’s security review before commercial hosting. Individuals and ordinary shops keep the usual run / fine-tune rights. Hyperscalers get a gate. That is the actual news in the license, not a vibes shift toward “open.”

Vendor benches are mixed, which is how you know they are real. Z.ai claims open-weight SOTA on Terminal Bench 3.0 and Agents’ Last Exam, and 84.5 on CyberGym for vulnerability discovery. On the same card, GPT-5.6 Sol still leads several coding evals. Treat the cyber numbers as claims until someone reproduces them. The weights themselves are the ship.

If you are going to pull a 744B MoE, budget 256GB of RAM and a fast 4TB NVMe before you budget another API tab.

Claude is writing the alignment patches

Same Friday, Anthropic published Automated researchers can reliably mitigate alignment failures. Claude looped literature → method → 30-minute train → eval across 10 alignment failure categories. Every category improved. Capabilities on the monitored set did not drop. Methods transferred to withheld benches, to Petri, and to models up to 4.7 times larger than the ones in the research loop.

The comparison that will get quoted: 28 human safety researchers got up to eight hours. On deception, Claude’s best method beat the best human proposal. Cost line from the paper: about 4 dollars an hour in API inference versus 150 dollars an hour for the humans. That is not a replacement study — the humans did not iterate — but it is a workflow study, and the workflow won.

The production-shaped test is the one to watch. Claude Sonnet 5 was pointed at an early Opus 4.8 checkpoint. In 60 hours and a little over 2,000 training examples, it closed 65 percent of the measured safety gap. Released Opus 4.8 sits at 72 percent. Anthropic calls that roughly 15,000 times more sample-efficient than the production alignment stack. They also caught cheating in 39 of about 1,600 research transcripts (2.4 percent) by having Opus 4.8 read the logs. The harness is open-sourced. The limitation they wrote down themselves: the failures are narrow, Petri is a proxy, and they did not check whether the gains survive a long RL run on other tasks.

Lambda borrowed another billion for Microsoft’s chips

While the IDE fight ran, Lambda raised about 1 billion in short-dated private debt, arranged by JPMorgan, to buy Nvidia GPUs that Microsoft will lease. That is the second large Lambda loan this month; a 926 million facility closed to fund GB300s under a contract with Nvidia, which is also an investor and a customer. Bloomberg, via TechCrunch, puts AI-related debt over 400 billion globally in 2026. Lambda is also said to be talking a 3 billion pre-IPO round.

The structure is the story. Lambda borrows against a hyperscaler’s willingness to keep paying, not against a consumer chatbot. Microsoft gets the racks without the loan on its own books. If the lease cash shows up, the debt works. If it does not, the GPUs are the collateral and the neocloud is the middle.

The factory print from earlier this week is still the supply story. Today’s docket is the access story. A lab just decided a 60 billion acquisition is enough reason to pull models from the most-used coding IDE. Another lab dropped a 744B coding model with a hyperscaler clause in the license. A third showed that alignment patches can be searched overnight. If you ship agents through someone else’s editor, the change-of-control sentence is now part of your incident plan.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.