A video company published research arguing that the way software gets built — write code, let a browser draw it — is an unnecessary detour, and showed a model drawing the screen directly instead. Elsewhere: OpenAI cut off one of the most popular coding tools in the world over who now owns it, started quietly testing a pricing model where customers pay only when the work succeeds, and crossed a billion dollars of annualized advertising revenue. Regulators sued Amazon over seven years of hidden ad surcharges, Nvidia spent on two fronts at once, and the most honest robotics reporting of the week was a long list of things humanoids still cannot do.
Runway's Solaris draws the interface instead of coding it
Runway introduced Solaris, which it calls an "Interface World Model." Every AI coding tool today writes code, and your browser turns that code into buttons, menus and screens. Solaris skips the middle step: it generates the interface itself, one frame at a time, predicting what the screen should look like after each click or drag. Runway describes treating user input as conditioning for the next frame, so the model learns the relationship between an action and its visual result without anyone programming that relationship. A language model decides how the interface should evolve; Solaris does the drawing.
The argument for doing it this way is a measured claim about information loss. Runway tested whether frontier models can faithfully reconstruct an interface from an image, and found every one of them — Claude and Gemini included — drops detail, with degradation rising as visual complexity rises. In a study of 250 evaluators across 30 interaction examples, judges preferred Solaris to interfaces coded by Claude Opus 5 in 61% of instruction-following comparisons against 24%, and 71% against 21% on whether elements behaved naturally within the scene.
Runway is candid about what does not work yet. Legible text remains, in its words, one of the hardest problems. Coherence degrades over long sessions, the model can produce confidently incorrect responses, and integration with screen readers and accessibility APIs is unresolved — which for any business with customers or employees who use assistive technology is not a footnote but a gating requirement. Solaris is request-only early access, not a product you can buy. Treat it as a signal about where interfaces are heading over several years, not something to plan around this quarter.
OpenAI cuts off Cursor, and the frontier picks sides
OpenAI will end Cursor's direct access to its models on November 12, after SpaceX's $60 billion acquisition of Cursor's parent company, Anysphere. OpenAI invoked what it described as a limited termination right in the agreement, saying it made the choice "because we cannot be confident that SpaceX will use our technology within our terms of service" — a reference to prior disputes with Musk-owned companies. Cursor CEO Michael Truell says OpenAI models serve only about 5% of the platform's traffic; the product also runs Anthropic's Claude models, Google's Gemini, and xAI's Grok. Musk's public response was that he could not care less.
Nobody's editor breaks on November 12. But the precedent matters more than the outage. A model vendor terminated a major customer over that customer's new ownership, using a clause the customer did not control, and the tool's users found out from a press cycle. That is the shape of the market now: frontier access is a relationship subject to whitelists and change-of-control terms, not a utility you buy by the unit. If a tool your business depends on has one model provider behind it, you have inherited a dependency you did not negotiate and cannot see.
Pay only when it works
OpenAI has started testing an arrangement with a small number of large accounts where they pay only when the AI actually completes the job, according to reporting by Kevin McLaughlin and Amir Efrati at The Information. The customers, terms and prices are all undisclosed, OpenAI has not announced the test, and the reporting has not been independently confirmed — so treat this as a credible report rather than an established product. The example cited is customer support handled end to end.
This addresses the single hardest thing about budgeting for AI: today's bills scale with attempts, not results. A finance director can defend an invoice that arrives when something worked far more easily than one that arrives regardless. Two vendors already sell this way, which is the useful part for smaller buyers. Intercom charges $0.99 per resolved conversation. Zendesk bills only for "Verified Resolutions" — outcomes confirmed by an automated evaluation within 72 hours — at roughly $1.20 to $1.50 on committed volume. Those are published prices you can benchmark a support-automation proposal against today, without waiting for OpenAI.
Meanwhile OpenAI's advertising business reached a $1 billion annualized revenue run rate roughly 200 days after launch, having crossed $100 million within six weeks of the US pilot in April. Ads now run in more than 40 countries, and the self-serve platform opened Monday to marketers in India, Europe, the Middle East and North Africa. Run rate extrapolates current revenue rather than counting money already banked — but the direction is unambiguous, and the assistant your customers ask for recommendations is becoming an ad-supported surface.
Consolidation, chips, and a 305-billion-parameter giveaway
Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion, first reported by The Information with Reuters matching the figure. Neither company has confirmed it, and separate reporting suggests negotiations may not have produced a signed agreement — so this is a report, not a done deal. If it closes, the default place developers go to find and download open models, hosting over 3 million of them, would belong to the company whose chips those models mostly run on.
Nvidia also invested $3.5 billion in Taiwan's MediaTek through convertible bonds, extending a partnership across data centers, PCs and cars. MediaTek will adopt Nvidia's NVLink Fusion platform, which lets cloud providers and model developers build custom accelerators that plug into Nvidia rack-scale systems — an attempt to keep in-house silicon efforts inside Nvidia's ecosystem rather than outside it.
Against that consolidation, DeepSeek published the weights for V4-Flash-Vision-Exp on Hugging Face under an MIT license on August 31, ten days after the same model appeared on its API. The repository metadata lists 305 billion parameters, with 284 billion in the mixture-of-experts backbone and the remainder covering the vision tower and supporting modules. DeepSeek says it matches its text-only V4-Flash on reasoning and world knowledge while closing much of the multimodal agent gap to Opus 4.8. Whatever else is true about the market's direction, a capable open multimodal model with a permissive license still lands every few weeks.
Regulators move on advertising, records and training data
The FTC and 22 state attorneys general sued Amazon on August 31, alleging that for more than seven years the company secretly inflated prices in its search advertising auctions. The specific mechanism alleged is a 2019 change to the auction rules that added an undisclosed surcharge Amazon internally called a "soft reserve price." The complaint says the scheme affected more than one million brands and sellers and likely extracted tens of billions of dollars from advertisers who did not know it existed. Regulators have sued Amazon before over its marketplace and Prime enrollment, but this is the first action aimed at the ads unit — the part of the business many small sellers treat as a fixed cost of visibility.
The European Commission designated ChatGPT, Reddit and Roblox under the Digital Services Act after each crossed 45 million monthly EU users. And in Sony's litigation against Anthropic, internal staff messages cited in filings allege mass torrenting of books beginning in July 2021, that co-founder Benjamin Mann personally downloaded and seeded pirated titles, and that Dario Amodei approved the practice. Anthropic denies using pirated material to train its commercial models; publishers say they can prove otherwise. These are allegations in an active case, not findings.
Agents got a big open-source release and a forecasting model
OpenClaw shipped 2.0, its largest update yet: more than 16,000 pull requests from 933 contributors — 569 of them first-timers — representing roughly half of every pull request ever merged into the project, after about seven weeks without a release. The most practical change is installation, which now reuses what is already on your machine, including existing ChatGPT or Claude subscriptions, API keys and local models.
Google released TimesFM-3, a 330-million-parameter time-series foundation model pretrained on more than a trillion time points. It forecasts multiple targets at once with no task-specific fine-tuning and accepts both historical and known-future variables — meaning a small business could forecast demand while telling the model about a planned promotion, without training anything. Elsewhere, the Department of War launched ChatGPT Mil on GenAI.mil for more than three million personnel, and John Ternus takes over as Apple CEO today after Tim Cook's fifteen-year run.
Physical AI
The most valuable robotics reporting this week was a careful accounting of what humanoids cannot do, assembled from interviews with the people building them. Start with the number that should anchor every humanoid conversation: when Physical Intelligence demonstrated ten of fifteen manipulation tasks from roboticist Benjie Holson's challenge list — opening doors, making a peanut butter sandwich, peeling an orange — the robot succeeded 52% of the time and took four to ten times longer than a person. Teaching it to turn a sock inside out required 176 human-puppeted demonstrations, about eight hours of data. These are tasks trivial for an eight-year-old. Holson released a harder set in January — make a bed, hammer a nail, catch an egg — and no one has announced solving any of them.
Endurance is the constraint nobody demos. Unitree's G1, one of the most widely deployed humanoids in the world, historically could carry a few kilograms for about five minutes before overheating and needing thirty to sixty minutes to cool, according to SemiAnalysis robotics lead Reyk Knuhtsen; after design improvements, operators now get five to fifteen minutes of work per ten minutes of rest. Agility Robotics keeps its deployed humanoids inside safety enclosures away from workers, and CTO Pras Velagapudi described chasing a field failure that turned out to be a circuit board in the leg flexing over thousands of steps until crouching disconnected a cable. 1X, which plans home deployments this year, told the Wall Street Journal that families with young children cannot join its testing program. In a Chinese pharmaceutical warehouse, a Galbot robot took roughly 40 seconds to move a single box from shelf to chute, and needed retraining — the company says about five minutes — every time a new item entered the catalog.
The showcase version looked different and ended badly. At the 2026 World Humanoid Robot Games in Beijing, a robot ran 100 meters in 8.86 seconds, comfortably under Usain Bolt's 9.58-second human record, then hit the stopping barrier, collapsed, and briefly caught fire around its torso; officials extinguished it. Robots across the event splintered, fell apart and were carried off. The gap between an 8.86-second sprint and a 52% success rate at peeling fruit is the whole story of this field right now.
Where robots are earning revenue, the jobs are narrow and the form factor usually is not humanoid. Gatik closed a $200 million Series D on August 25 led by the Qatar Investment Authority and Koch Disruptive Technologies, with Millennium Management and ARK Invest participating, to expand driverless middle-mile freight for large retail, grocery and consumer-goods supply chains — fixed, repeated routes rather than open-ended driving. At the other end of the price range, Pollen Robotics is selling Microduck, a small biped with fifteen servos, for $399; the company says it has sold more than 10,000 units. For an operator, that is the practical read on this beat: the deployable systems are the ones doing one repetitive thing on a known route, and the cheap end of the hardware is now genuinely cheap enough to experiment with.
Quick Takes
Amplitude says its Wave agent chose and shipped its own product changes, moving key metrics by 50% to 3× — a vendor claim about its own product, not independent measurement, but an early example of an agent given authority over a live product surface.
Pieter Levels launched Infinite Slop, an interactive AI livestream where viewers steer what happens next, including an automated news segment built from X and Hacker News. He estimates it costs about $4,000 a day to run without sponsorship — a useful number for anyone pricing continuous generative video.
Fireworks made its Training API and Lab generally available, letting teams fine-tune open models for specific jobs and deploy them from the same platform.
Meta's Muse Code shipped, a terminal and CI coding agent with approvals and an OS sandbox on by default, alongside Muse Spark on the Meta Model API.
Instagram now requires profiles built around AI-generated personas to carry an "AI-generated profile" label, replacing the previous optional disclosure.
Microsoft spent Monday managing a broad Microsoft 365 outage that began with Exchange Online and spread to Teams, OneDrive and SharePoint.
What This Means for Your Business
Ask every AI vendor you rely on which models sit underneath their product, and what happens if one gets cut off. The Cursor termination is the cleanest illustration yet that model access is a contractual relationship between two companies you are not party to. You do not need to panic-diversify; you need to know the answer. For each AI tool your team uses daily, write down the underlying model provider and whether the vendor can fail over to another. Tools that route across Anthropic, Google, OpenAI and open models are structurally safer than tools wired to exactly one, and that is now a purchasing criterion rather than a technical detail.
Price your next AI project on outcomes, and use the published numbers as leverage. You do not have to wait for OpenAI's experiment to reach you. Intercom's $0.99 per resolved conversation and Zendesk's roughly $1.20 to $1.50 per verified resolution are public benchmarks you can hold up against any support-automation quote. When a vendor proposes usage-based pricing, ask what a resolved outcome costs at your volume, and ask who decides whether it was resolved and within what window. The 72-hour verification window in Zendesk's model is the part most proposals leave vague, and it is where the disagreements happen.
If you advertise on Amazon, the FTC complaint is a prompt to audit, not just to read. The allegation is that an undisclosed surcharge inflated auction prices for over seven years across more than a million sellers. Whatever the case's outcome, pull your advertising cost data back as far as your account allows and look at cost-per-click trends against your own bid changes. If there are step changes you cannot explain by your own actions or by seasonality, document them now while records are easy to retrieve. More generally: any channel where you bid into an auction you cannot inspect deserves a spend cap and a quarterly review.
On robotics, buy the boring thing. The honest reporting this week points one direction — the systems generating revenue do one repetitive task on a known route or a fixed station, and the impressive ones are demos with 52% success rates and ten-minute duty cycles. If automation is on your roadmap, look for the single highest-volume, most repetitive, most physically consistent task in your operation and price automating just that. Ask any vendor two questions the demo will not answer: how many minutes of continuous work before it needs to rest, and how much retraining a new product or SKU requires. Galbot's five minutes of retraining per new item is a good benchmark to hold others to.
Finally, decide now whether AI-generated content on your properties gets labeled, because the platforms are deciding for you. Instagram's mandatory AI-persona label and the EU's designation of ChatGPT, Reddit and Roblox under its strictest rules are the same trend: disclosure is moving from optional to enforced. If you use AI to generate images, video, personas or reviews in your marketing, write down your labeling policy this month rather than discovering a platform's version of it after a takedown. The cost of disclosing voluntarily is small; the cost of being caught not disclosing is a distribution problem you cannot appeal.