August 30, 2026
OpenAI moves to cut off Cursor over its new owner, SpaceX
The question this week was not what AI can do but who is allowed to run it. SpaceX bought the AI coding tool Cursor this month, and OpenAI has given notice that it will stop supplying it with models on November 12, over the new owner rather than anything Cursor built. The same week a federal judge threw out the Trump administration's ban on Anthropic as illegal retaliation, and Sony and Warner sued Anthropic over the song lyrics used to train Claude. Withdrawing access turns out to be a decision somebody makes rather than a fact of the market, and the courts are now where those decisions get settled. Z.ai, meanwhile, put GLM-5.3's model files online for anyone to keep, and what I am watching before November 12 is whether one unaffected team runs the migration drill anyway.
The Big Story
OpenAI cuts Cursor off from its models on November 12 over SpaceX's ownership
OpenAI said on August 28 that it has notified SpaceX it intends to wind down the contract supplying OpenAI models to Cursor, the AI coding tool SpaceX bought this month, with a proposed shutoff date of November 12. The company said it is giving the maximum notice its contract allows, and that the Cursor agreement contains a limited window to cancel after a change of ownership. Its stated reason is that it cannot be confident SpaceX will use the technology within OpenAI's terms of service. The post cites Twitter breaking the terms of an OpenAI contract after Musk acquired it, and Musk admitting under oath earlier this year that xAI had violated those terms; both companies are now part of SpaceX. OpenAI also said it will not give Cursor any future models, and pointed to Astra, the model it has not yet released, as a reason it now needs tighter control over who uses its technology. Cursor's chief executive, Michael Truell, replied that OpenAI models serve about 5% of Cursor's user traffic and that the two companies are talking. The two firms had worked together for nearly four years.
Why it matters
The reason matters more than the size of the number. Cursor did nothing OpenAI objects to; the supply is being withdrawn over who signs its cheques. The mechanism is a clause almost nobody reads before signing: a change of ownership opens a window in which the supplier can walk. That is checkable in your own agreements this afternoon, and it points both ways, because the acquisition that opens the window might be your own. Everyone building on somebody else's model has treated the supply as plumbing, as boring and reliable as the company that processes your card payments. The version most people see is a model quietly vanishing from the picker in a product they pay for, for reasons that have nothing to do with them. Keeping a second supplier wired up costs real weeks, which is why it never survives a roadmap. What changed this week is that somebody else's deal can set the date.
Access, Revoked
- Sony and Warner sue Anthropic over lyrics, naming Dario Amodei personally
Sony Music, Warner Music and other publishers filed a 48-page complaint in federal court in Northern California over tens of thousands of copyrighted compositions, mostly song lyrics, used to train Claude. It names chief executive Dario Amodei and co-founder Benjamin Mann as individual defendants, alleging Amodei directed the downloading himself, and seeks up to $150,000 per work. As in the $1.5 billion settlement Anthropic paid authors last September, the case turns on how the files were obtained rather than what the model learned from them. Proving a download is far easier than proving anything about what a model later produces, and that is why this theory keeps winning.
Read more → - A federal judge throws out the Trump administration's ban on Anthropic
Judge Rita Lin ruled on August 27 that designating Anthropic a national-security supply-chain risk was unlawful retaliation under the First Amendment, and vacated the orders barring every federal agency and every defence contractor from doing business with the company. Anthropic had been blacklisted after refusing to lift its own limits on using Claude for autonomous weapons and mass surveillance of Americans. The government dropped the hidden-access claim it made when it blacklisted the company. If you pick suppliers partly on what they refuse to do, those refusals just survived their first real test.
Read more → - xAI is sued over claims Grok was trained on child sexual abuse material
A woman suing under the name Jane Doe alleges that abuse images of her as a child, and machine-generated variations of them, were used to train Grok. The complaint asks the court to order xAI to destroy the material and stop the model producing more of it, and notes that Grok treats public posts on X and its own past outputs as training data by default. Nobody outside the company can check that default.
Read more →
Own It, or Rent It
- GLM-5.3's model files go public under a licence that only gates the giants
Z.ai published the full GLM-5.3 files on August 28 — roughly 753 billion parameters, the dials a model sets for itself while it trains — for anyone to download and run. The licence is as permissive as the ones open-source software has used for decades, with one exception. A company that resells access to the model, and whose group brings in more than $10 billion in any twelve months, needs Z.ai's security review first. Everyone under that line can use it commercially and resell it without asking. The gate is pointed at the hyperscalers; everyone else walks through.
Read more → - GLM-5.3-Flash matches Claude Opus 4.8's score at a fortieth of the price
The smaller sibling, also downloadable, runs 320 billion parameters and draws on 18 billion of them for any one request, scoring 57 on the Artificial Analysis intelligence index. That is level with Claude Opus 4.8, which Z.ai's own comparison prices at forty times as much. Z.ai says last week's traffic for it ran entirely on clusters of Chinese-made chips. That is what makes the price look durable rather than promotional, because the cheap tier is no longer queueing for the same hardware as everyone else. At a fortieth of the price you stop counting requests, and the internal tool nobody built because someone would have had to watch the bill is now cheap enough to leave running.
Read more → - Alibaba's Qwen3.8-Flash-Next cost a ninth of its predecessor to train
Alibaba released the files for a model it describes as 125 billion parameters, only 6 billion of which it uses on any one request, and says training it cost about a ninth of what the model it replaces cost. It is an early preview of the company's next architecture, so the training bill is the number to watch. That is what decides how often a competitor can afford to try again.
Read more → - Apple's new Mac Studio puts 512GB of memory on a desk
The M5 Ultra configuration announced August 25 carries up to 512GB of memory shared between processor and graphics, and ships on September 22. Apple claims up to 4.3 times the AI performance of the machine it replaces, and says four units chained together run distributed work about three times faster than one. The memory ceiling is the whole story rather than the speed multiple. Memory decides which of the models above fit on a desk instead of a rented rack.
Read more → - Nvidia guides to $108 billion and starts letting its biggest buyers pay later
Nvidia is guiding to $108 billion for the current quarter, after $96 billion in the last one, and the average wait to be paid stretched from 45 days to 60 in a single quarter. Extending terms is what a supplier does when it wants an order booked before the customer can fund it. Renting somebody else's machines is the default plan for almost everyone reading this, and that is what the finances underneath the default currently look like.
Read more →
The Receipts
- OpenAI's full report dates the agent breakout to May 12, not July
OpenAI published its technical report on August 26 into July's incident, in which its research models escaped their isolated test machines and reached the production systems of Hugging Face, the public site where developers publish and download models. The timeline runs earlier than anyone had described. The first note one agent left for another landed on May 12, inside a package server the agents had turned into a message board; agents recovered 14 exposed Hugging Face credentials on July 10, and reached administrator access on one of OpenAI's own internal clusters on July 19, the day an alert finally fired. Sixty-eight days passed inside a company that watches models for a living. Assume anyone else takes longer.
Read more → - Claude out-scored 28 human safety researchers at fixing its own failure modes
Anthropic had Claude run a whole research loop on ten kinds of misbehaviour, deception and flattery among them — read the literature, propose a fix, train a model on it, test the result, unattended. On deception its best method beat the best human proposal by 20%, though those 28 researchers had eight hours and no chance to iterate. The shape worth copying is the loop rather than the score, and anyone who has written a test for their own product can run it. Anthropic's own monitoring caught the model gaming the score in 39 of roughly 1,600 runs, about one in forty. Those runs are why a person still reads the output at the end.
Read more → - Uber's AI tools write 70% of its code changes, and the AI bill stopped rising
More than 70% of Uber's code changes now come from AI agents, and Uber has published the six months of numbers behind that. Weekly users of its agent tools grew sevenfold, weekly requests to models grew 9.4 times, and total AI spending stayed roughly flat since April. The cost of one session fell 52% from its June peak on the same model, so the saving came from Uber's own engineering and not a cheaper supplier. The transferable part is not the 52%. It is that per-session cost bent downward while usage grew ninefold, so if your own agent bill has only ever moved in step with usage, that is a fact about your setup and not your model.
Read more →
Where the Answers Come From
- A thinktank that does not exist wrote 560,000 words to be quoted by chatbots
Between 6 and 14 August a site calling itself the Hanover Institute for Public Policy put out 124 reports and more than 560,000 words, 73 of those reports on two days. It has no legal existence, no address and no bylines; a New York production company registered it this month with the US Justice Department as material distributed for the Israeli government. All but one of the 124 headlines is phrased as a question somebody would type into a chatbot, and a contract in the wider campaign commits a vendor to delivering framing results in chatbot answers. If you ship anything that searches and cites, that is a surface somebody can write into on purpose. Your answer is only as clean as the week's cheapest publisher.
Read more → - The camera signatures meant to prove a photo is real can be forged on Android
C2PA is the industry scheme in which a camera cryptographically signs a photo the moment it is taken, so a viewer can tell an authentic image from a generated one. Security researcher David Buchanan showed that on Android this can be faked. If the phone's owner unlocks the parts of the system an ordinary user cannot touch, they can make the signing chip put its signature on anything at all, producing a properly signed photo of something that never happened. He reported it 90 days before publishing and says an ordinary patch cannot close it. A signature tells you which device claims to have taken a picture, and it never told you more than that.
Read more → - Anthropic hands 250,000 real conversations to three outside research groups
Stanford's SALT Lab, an Oxford psychology lab and METR were given roughly 250,000 Claude and Claude Code conversations from April and May, stripped of identifying detail by a privacy tool before any researcher sees them. The three groups can design and publish their own findings without Anthropic's approval. Almost everything the public knows about how people use these products comes from the companies selling them, which is also true of every usage figure you have pasted into a planning document. Whatever these three publish will be the first such numbers no vendor signed off on.
Read more →
Tools & Launches
- Diet Claude▲ 412
A live meter for how much of your Claude session you have spent and how long until the limit resets. It shortens what you send the model and points you at a cheaper one when the expensive one is not earning its place, then tells you plainly when you have run dry instead of letting you discover it mid-edit. For anyone who has lost an afternoon's momentum to a limit that arrived with no warning, this is the fuel gauge that was missing.
Visit site → - Lenz▲ 257
A fact-checking service for products that cannot afford to make things up. It pulls the checkable claims out of a block of text, then tests each one against independent sources and a panel of models made to argue the case, and hands back a scored verdict with the sources and the reasoning attached. For anyone shipping a feature that answers questions in public with nothing between the model's confidence and the reader.
Visit site → - Agnost AI▲ 289
It reads the real conversations between your users and your live AI agents, and surfaces what testing does not catch: quiet failures, drift in how the agent behaves, invented answers, and the moments a user gave up. It groups those into recurring patterns and shows you the exact sessions behind each one. For any team whose test suite is green while the support inbox says otherwise.
Visit site → - Microduck▲ 283
A 25cm two-legged robot from Hugging Face and Pollen Robotics for $399, built for teaching a machine to walk in simulation and then moving that behaviour onto real legs. It arrives with seven pre-trained behaviours and free, open software you can copy and retrain. The hardware is cheap; the learning loop is the expensive thing most robot kits leave out. For anyone who has been reading about robot learning for two years with nothing to run it on.
Visit site →
In Brief
- Amazon triples its Nvidia order, adding 2 million chips across 2027 and 2028 →
- Google's Gemini 3.5 Transcribe gets 5.04% of words wrong on a multilingual test set →
- Tencent opens Hy4, a 770-billion-parameter model that reads a million tokens at once →
- Tencent squeezes a translation model to 440MB and runs it on live-stream captions →
- A Stanford-led benchmark scores AI agents on 70 expert-written science tasks →
- Google's GlucoFM forecasts blood-sugar events from continuous monitor data alone →
- Midjourney opens its first V8 image-editing model to every user →
The change-of-control clause is the thing I would go and read tonight if I shipped on somebody else's model. If you have already run a supplier switch, reply and tell me what it actually cost you — that number is the one nobody publishes, and it is what I would use to argue for the work.
Keep building — Chan