What The Top AI Users Are Doing Differently

August 25, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
August 25, 2026

HEADLINES

Meta Readies Hatch

According to roadmap documents viewed by The Information, Meta is putting the finishing touches on its consumer agent ahead of a release in the coming weeks. Internally known as Hatch, it sounds like it sits in the GrokBot family of delivering a more streamlined version of an OpenClaw-style agent experience, and could anchor a new AI agent subscription with a $200 a month price tag for high usage accounts. A new WhatsApp platform will let third-party agents coordinate with each other over messages — not Meta cribbing off GrokBot so much as an interaction pattern everyone is converging on. Also coming: a larger model called Watermelon, targeted at October, which Alexandr Wang told staff in July had already caught GPT-5.5 on internal benchmarks. The frontier has moved a long way since then, so the open question is whether Watermelon keeps pace or lands in the familiar column of Meta getting closer without arriving.

GrokBot and Sol Get Cheaper

One of the barriers to testing GrokBot was pricing that was not just extreme but confusing — users initially weren't sure whether they needed both Cursor Ultra and SuperGrok Heavy at a combined $500 a month. That was always an intentionally rate-limiting launch to keep things from falling over, and as of this week GrokBot is included in the $60 a month Cursor Pro subscription and the $100 a month SuperGrok subscription. OpenAI cut prices in the same week, with GPT-5.6 Sol dropping to $4 per million input and $20 per million output tokens, down from $5 and $30. The popular read is pressure on Anthropic ahead of the IPO, but the likelier driver is OpenAI recognizing that business customers are no longer running the most expensive state-of-the-art model for everything and are thinking in terms of a complete model stack — so to the extent the compute exists to deliver frontier models cheaper, it makes sense to do it.

Hugging Face Revenue Up 50% in Two Months

Following yesterday's report that Hugging Face is courting an acquisition at a $13B valuation, sources speaking with The Information say the company is now generating more than $150M in annualized revenue, up 50% from two months ago. That number looks low next to the coding agent startups, but Hugging Face has specifically not been focused on generating revenue — 97% of users access the platform entirely for free, including downloading the latest model weights, with premium and enterprise accounts, inference and hyperscaler partnerships existing largely to subsidize the free hosting and distribution of open models. Clem Delangue said in June that premium accounts had doubled in the first half and that profitability was close. For acquirers this is very likely not a strict revenue multiple conversation, which is the logic Jess Fields sums up in arguing Hugging Face should be worth what Cursor is, three to four times the asking price, because it is the backbone of open weights challenging frontier gatekeeping.

What Is Jensen Building?

With a flurry of reporting around NVIDIA's dealmaking, some are wondering what Jensen is actually assembling. In the past week alone: a deal to license technology and acquire talent from Poolside, a stake in data labelling company Mercor, and a potential investment in Perplexity at a perplexing $30B valuation — to say nothing of equity in neoclouds, land and power deals and data center backstops. Some have started conceptualizing NVIDIA as the central bank of compute, while Martin Peers compares Jensen to John Malone's 1980s cable empire; another rough comparison is Alphabet's Other Bets. Before screaming circular dealmaking, the thing to recognize is that NVIDIA is more or less tapped out on reinvesting in its own business — it doesn't operate fabs, so chip revenue growth is constrained by suppliers it can't control. Turning investment outward supports the entire AI economy, which is what keeps its own revenues strong for years.

Taiwan Charges Nine Over Blackwell Smuggling

Taiwanese prosecutors announced nine indictments on Monday over a scheme to smuggle cutting-edge Blackwell 300 systems into China, including a manager in NVIDIA's distribution business and two people who worked for NVIDIA partner Supermicro. A Supermicro co-founder was charged earlier this year over a separate incident, and both companies have indicated the problems are contained to a few rogue employees and that they are working with authorities. To give a sense of scale, the group allegedly ordered 130 Supermicro servers containing Blackwell 300s, with prosecutors claiming 74 were delivered to buyers in China and another 56 stopped by Taiwanese officials — in total less than 10,000 chips, which is not nothing but certainly not enough to build a frontier training cluster.

Inside Microsoft's AI Spend

Business Insider got hold of an internal spreadsheet where Microsoft employees self-report salary, bonuses and AI usage. This is not a story about token maxing — BI found no correlation between token burn and compensation or promotions — but it is an interesting peek at how jagged usage is both across and within departments. The Azure range ran from $1 to $7,500 a month, cloud and AI topped out at $15,000, and at least one person in Microsoft Customer and Partner Solutions ran up a $28,000 bill, while medians were far more clustered — seven of the eight departments around $150 to $500, with Core AI the outlier at $975. All of it is voluntary self-reporting from just 600 US employees, only 350 of whom reported AI usage, out of a 223,000-person workforce, but it still gives a sense of the magnitude among the more prominent AI users inside a company like Microsoft.

MAIN STORY

What the Top AI Users Are Doing Differently

There has always been a gap between the average AI user and the most advanced ones, but that gap has now exploded. Newly released OpenAI research shows the distance between frontier firms and typical firms widening from 2.6X in January to 8.3X by the end of June, and the reason is agents: as agentic use cases became viable this year they upgraded the difficulty, complexity and importance of the work AI could take on, and the top users jumped in headfirst while everyone else didn't. The compounding is the whole story — agentic use is not a model question but a question of how those models get put to work, and as power users hand agents increasingly valuable work, the gap is poised to keep growing.

Sam Altman on Being Wrong
"The economy just has so much inertia"
Altman has increasingly gone out of his way not just to change his position on rapid job disruption but to explain why he thinks he was wrong: he expected software businesses to be up for grabs right away after GPT-4 and they weren't, because people keep doing the same things and buying from the same companies. He frames the slowness as a positive that will make the transition go smoother, and says he's grateful for it. Or, put another way — who needs a pause AI movement when you've got corporations?

Revealed Preference
"By revealed preference, I have a better way and still do it the old way"
The more interesting half of the interview is Altman applying the inertia argument to himself: he has a magic thing called Codex and still clicks around, pastes between messaging apps and scrolls his inbox looking for the least painful email. He calls it psychologically inconsistent and concludes he must secretly like it. It's probably not secret satisfaction — it's that everyone builds intellectual mind muscle memory, and undoing it to build a new and much less comfortable one looks exhausting when the old way will get it done right now.

Legal Is Up 108X
Non-technical roles are growing agentic usage fastest by a mile
Indexed to the beginning of February, every role has gone up, but the non-technical ones have gone up hardest — partly a lower starting point, but the spread is striking. Engineering and technical practitioners are up 5X, finance and accounting 20X, marketing and communications 26X, people and recruiting and sales and account management both 41X, and legal a full 108X. On the lawyer side of this, Spellbook's Scott Stevenson quotes Jack Newton: LLMs are for lawyers what spreadsheets were for accountants.

Assistance to Delegation
Not a model question — a question of how models get put to work
The through line of the report is two-part: more work is being delegated to agents, and because of that there's a compounding effect where the firms farthest along keep pulling away. OpenAI's own framing is that the difference is moving from assistance to delegation, giving agents the context and tools to complete complex tasks, and pushing agentic use beyond software development.

The Flippening
64% of enterprise output tokens are now agentic, from roughly zero
A year ago the balance between ChatGPT and agentic usage, measured by enterprise output tokens, was basically 100% ChatGPT. Codex went generally available in October and pushed agentic into low single digits; by early February it was 87/13, by March 73/27, and by late April the flippening arrived at 53% agentic. By June, where the data ends, it's 36% ChatGPT to 64% agentic. That doesn't mean two-thirds of sit-downs at work involve agents — output tokens are a proxy for volume of work, and the preponderance of the work is now agentic.

The 8.3X Gap
Frontier firms are using 17 times the tokens they were 18 months ago
OpenAI defines frontier firms as the top 10% of usage in a month by output tokens per active user, with typical firms between the 45th and 55th percentile. Through 2025 the gap sat at roughly 2X, started widening in October, stood at 2.6X in January and has now absolutely exploded to 8.3X. The average firm is using about twice as many tokens as a year and a half ago; frontier firms are using 17 times as many.

Skills and Plugins
Six times as many people using skills, and OpenAI is at 93%
Part of the gap is using agents more and part of it is being better at using them, which OpenAI measures through weekly active users touching plugins or skills. At typical firms about 9% use plugins and only 3% use skills; at frontier firms it's 21% and 19% — roughly two and a third times as many people on plugins and more than six times as many on skills. Frontier firms haven't topped out either: inside OpenAI, 93% of employees use skills and 95% use plugins.

Why Knowledge Work Lagged
Software moved first because code bases give agents clear context
OpenAI's diagnosis is that software moved first for a reason — code bases give agents clear context, tests make outputs easy to verify, and coding progress accelerates AI R&D itself — while general knowledge work lagged because tasks provide limited context, are hard to specify and lack clear criteria for verifying results. Their explanation for the change is scaling, RL advances and targeted work on evaluations like GDPVal. It is great that OpenAI is working hard on this, but what OpenAI has done is not the reason agentic use has grown among non-engineering knowledge workers. The reason is that users have started to figure out the patterns that let agents thrive in their own contexts.

The Use Case Ladder
Generation, synthesis, execution, maintenance — chat stops at the bottom rungs
There's a massive shift in the type of work when you move from chat to agentic, and the ladder is the way to see it: generation at the base (drafting an email, a report, an Excel formula), then synthesis across disparate sources, then execution inside existing systems, then maintenance of a system over time. In legal, chat is 57% writing and 20.5% knowledge retrieval; move to agentic and retrieval falls to 8.3% while system operations jumps to 17.7%, workflow automation to 7.7%, classification and extraction to 4.6%, and coding — actual application building by people who are not software engineers — to 32.9%. The resulting division of labor has agents handling coverage and coordination while people still own risk judgment and accountability.

Multiplayer AI Is Next
The next gains come at the intersection of different teams
Even the frontier users climbing into execution and maintenance are still mostly running agents inside individual silos. The thing likely to supercharge this to the next level is the emergence of multiplayer and team AI rather than individual AI, because that's where the next generation of gains lives. None of this is all that surprising — it's been clear for a while that 2026 was the year agents became real — but the numbers do not lie, and frontier firms racing ahead while barely beginning to figure it out should be a wake-up call for anyone not deploying agentic use at scale yet.