A coding tool can lose access to a major model because its owner changed. That is the sharpest tension in AI news today: OpenAI says it will stop supplying models to Cursor after the coding platform’s acquisition by SpaceX, with a proposed cutoff on November 12. For developers, this is less celebrity feud than architecture lesson. Model access is now a dependency that can move with ownership, contracts, and policy—not only benchmarks or price. Friday also brought a 770-billion-parameter open model from Tencent, new compute accounting inside Gemini Notebook, and South Korea’s plan to make domestic AI broadly available as a public service. Meanwhile, U.S. officials corrected an important cyber claim, and California’s camera network showed what narrow, human-supervised AI can do well. The common thread is control: who supplies the intelligence, who pays for it, and what happens when the terms change while your workflow is already running.

AI news today: OpenAI gives Cursor a November cutoff

OpenAI said on August 28 that it notified SpaceX of its intention to wind down the custom contract that puts OpenAI models inside Cursor. The proposed shutoff date is November 12, the latest date allowed under the agreement, according to OpenAI. The company also says Cursor will not receive future OpenAI models during the transition.

OpenAI’s stated reason is that it cannot trust SpaceX to follow its terms, pointing to earlier disputes involving Elon Musk’s companies. That is OpenAI’s account; the contract itself is not public, and Cursor may still negotiate. The confirmed operational fact is that a change-of-control clause has become a product risk for developers. If your team standardizes on an AI coding layer, document which models are essential, what can be swapped, and how quickly prompts, rules, evaluations, and usage data can move. Our analysis of the OpenAI–Anthropic enterprise data divide offers useful context: provider choice increasingly bundles capability with governance and commercial trust.

Tencent opens Hy4, a 770-billion-parameter model

Tencent released the weights for Hy4 preview on Hugging Face under the Apache 2.0 license. The mixture-of-experts model has 770 billion total parameters but activates 49 billion per token, with a one-million-token context window. Tencent positions it for software engineering, document analysis, game development, and scientific research.

The performance figures are Tencent’s own. Its internal experts rated Hy4 slightly ahead of GLM 5.3 and Kimi K3 on 203 engineering tasks, but those evaluations are not a substitute for independent testing. Tencent also lists known problems: the preview can reason longer than necessary and over-verify its work. The consequential part is availability. Open weights, standard serving recipes, and an OpenAI-compatible API give advanced teams another model they can host and inspect—if they can afford the substantial infrastructure. “Open” lowers contractual dependence; it does not make 770 billion parameters fit under a desk.

Gemini Notebook turns usage limits into a compute budget

Google is changing Gemini Notebook’s consumer usage limits from simple daily allowances to compute-specific accounting. Limits will refresh every five hours and will vary with prompt complexity, conversation length, source count, and the feature used. When a Video Overview or Slide Deck exceeds the available budget, users can defer it for automatic generation later and request a notification.

The rollout starts September 2 on consumer web and mobile accounts. Google has not published a universal conversion table, so users should not assume identical quotas across tasks. The broader signal matters for every AI product leader: “one request” is becoming a poor unit of value when one request summarizes a page and another synthesizes a library into a video. Expect more interfaces to expose compute choices rather than hiding every cost behind a message counter.

South Korea treats AI access more like public infrastructure

South Korea selected consortia led by SK Telecom, KT, and Kakao for its government-backed “AI for All” program, Yonhap reported from the science ministry. The three groups will share 512 Nvidia B200 GPUs, with agreements expected in September and services targeted for launch by year-end. At least half of each system must use domestic models that meet government standards.

This is a reported implementation plan, not a finished nationwide service. Reliability, privacy, eligibility, and operating cost remain open questions. Still, the model is notable: instead of merely subsidizing labs, Seoul is using public compute and procurement to push homegrown models into citizen-facing services. That could improve access and strengthen local providers, but it also places unusually high pressure on evaluation and appeals when AI is used around health, taxes, benefits, or official procedures.

U.S. officials narrow the scope of a Chinese hacking claim

The U.S. Justice Department corrected an August 26 statement about a Chinese state-sponsored hacking campaign. The Senate, Federal Reserve, NASA, and other agencies were targets, but not all were compromised, Reuters reported after comparing the revised release with the FBI affidavit. NASA’s attempted breach failed because the targeted software had been patched. The affidavit does allege successful 2024 intrusions at three Energy Department laboratories, the NIH, an HHS agency, and a security-device manufacturer.

The correction is not a reason to dismiss the campaign; it is a reason to preserve distinctions. “Targeted” describes intent. “Compromised” requires evidence that defenses failed. Security leaders should insist on that vocabulary in incident reporting, especially as AI speeds up both attack analysis and rumor propagation. The boring discipline of patching also earned a rare starring role: in NASA’s case, it appears to have converted a target into an unsuccessful attempt.

California’s wildfire AI watches seven million images a day

ALERTCalifornia’s network of nearly 1,300 cameras now processes more than seven million images daily to look for early smoke, according to a Microsoft feature on the project. The team says the system detected 77 incidents before they were reported during its first two months and has spotted a fire as much as 2.5 hours before the first 911 call. Microsoft recently committed $5 million in technology development and Azure support.

Those results and counterfactual estimates come from project participants, so they should be treated as program claims rather than an independent effectiveness study. The design is nevertheless instructive: AI filters millions of frames, humans verify smoke and context, and firefighters make the response decision. This is not general intelligence replacing expertise. It is specialized perception buying experts minutes—the kind of measurable advantage many enterprise AI projects should envy.

Watch & Learn

Editor’s note: Google Cloud Tech’s “Context engineering explained: What every AI developer should know” is a practical 29-minute lesson on assembling instructions, data, memory, and tools around a model. It is best for builders and technical leaders who have moved beyond clever prompts and need repeatable agent behavior in production.

AI, Translated: context engineering

Context engineering is the work of deciding what information an AI receives, when it receives it, and how that information is structured. A support agent might get the customer’s plan, the current ticket, approved policies, and a tool for checking order status—but not the entire company drive. Prompt engineering improves the instruction; context engineering designs the whole information environment around it. You should care because reliable AI usually depends less on a magical sentence and more on supplying the right evidence and permissions at the right moment.

Try This Today: build a model-dependency audit in Claude

Goal: find out how painful a model cutoff would be before one happens. Claude Projects keeps related chats and reference material together; Anthropic’s official seven-minute introduction shows the current workflow. Availability can depend on your plan.

1. Create a Project called “Model Dependency Audit.”
2. Add a short architecture note and a sanitized month of model-usage data.
3. Ask Claude to separate portable components from provider-specific ones, then rank migration blockers by urgency.

Copy-ready prompt: “Audit this AI workflow for provider lock-in. Map every model, API, proprietary feature, prompt format, evaluation, data-retention rule, and contract dependency. For each, propose one substitute and one test proving the substitute works. Finish with a 30-, 60-, and 90-day exit plan. State assumptions and do not invent contract terms.” In ten minutes, you will have a better question than “Could we switch?” You will have a list of what switching actually means.

One thing to remember

The model is only one layer of an AI product. Contracts, compute budgets, context, public policy, and human verification decide whether that intelligence remains available—and useful—when conditions change.