OpenAI's own data says the gap between the most advanced AI users and everyone else has ballooned from 2.6X in January to 8.3X by June — and the whole difference is agents. Today's episode digs into what frontier firms are actually doing differently, plus Meta's imminent consumer agent, NVIDIA as the central bank of compute, and a peek at Microsoft employees' AI bills.
Agentic use compounds: the top AI users are pulling away from everyone else.
In January, the most advanced AI users were using 2.6 times as much AI as average users. By the end of June, per OpenAI's own research, that gap had grown to over eight times. The reason is agents: frontier firms have moved from chat-based assistance to delegation — giving agents context, tools, plugins, and skills to do systems-level work — and because agents take on increasingly valuable work, their lead compounds. The average users just haven't made the jump, and the gap is poised to keep growing.
One of the best shifts in AI narratives right now is moving away from the accepted-without-question premise that AI is obviously job-destroying. Certain roles will genuinely be obviated and society should be ready to support the people affected — but the idea of a radical, rapid jobs apocalypse was never accurate.
— Sam Altman, in a recent podcast interview. Altman now says he was wrong about the speed of disruption after GPT-4 — people keep buying from the same companies and using tools the same way. He calls the inertia a positive that will make the transition 'smoother and slower,' and says he's grateful for it.
Institutional inertia is doing the work AI doomers wished regulation would: even with incredible technology, society and the economy adapt slowly, and that slowness is the real governor on the pace of transformation.
— Sam Altman, on his own 'psychologically inconsistent' work habits. Altman admits he's been using computers the same way for 20 years despite having 'a magic thing called Codex' — still clicking around, pasting between messaging apps, scrolling mindlessly through email. Something in his mind encodes that rote computer work is what it means to be productive, even though his stated preference is the opposite.
We all get stuck in the patterns of how we've always done things. Building a new, less comfortable mind muscle feels exhausting when we know exactly how long the old way takes — which is precisely why studying the people who have broken out of that muscle memory matters so much.
OpenAI's 'Enterprise Signals: What Frontier Firms Are Doing Differently' puts hard numbers around how leading firms use AI — and the company barely promoted it. It only surfaced widely when A16Z reposted it in their charts of the week. Great stuff; please promote it more heavily next time.
Since early February, Codex use among engineering practitioners is up 5X — but finance and accounting is up 20X, marketing and communications 26X, people/recruiting and sales each 41X, and legal a staggering 108X. As one line making the rounds put it: LLMs are for lawyers what spreadsheets were for accountants.
In August 2025, enterprise output tokens were essentially 100% ChatGPT. By February the split was 87/13, by March 73/27, and by late April agentic crossed 53%. When OpenAI's data set ends in June, agentic work represents 64% of enterprise output tokens — meaning, using tokens as a proxy for volume of work, almost two-thirds of the work is now agentic.
Through most of 2025, the average user at a frontier firm (top 10% of enterprises by output tokens per active user) used about twice as many tokens as one at a typical firm. The gap hit 2.6X in January and now stands at 8.3X. Overall, average firms use about twice as many tokens as 18 months ago — frontier firms use 17 times as many.
At typical firms, about 9% of weekly active users use plugins and only 3% use skills; at frontier firms it's 21% and 19%. And even they haven't topped out: at OpenAI itself, 93% of employees use skills and 95% use plugins.
OpenAI's diagnosis: code bases give agents clear context, tests make outputs verifiable, and progress in coding accelerates AI R&D itself. General knowledge work lagged because tasks provide limited context and lack clear verification criteria — but scaling, reinforcement learning, and evals like GDPVal are bringing real-world tasks within reach, and agentic AI has found product-market fit with general knowledge workers since the start of the year.
It's great that OpenAI is making models and harnesses work better for knowledge work, but that's not why agentic use has grown among non-engineering knowledge workers. It's grown because we've started figuring out the patterns that actually allow agents to thrive in our own contexts.
Advanced agentic use climbs a ladder: generation (drafting emails, reports, formulas) at the base, then synthesis of disparate data sources, then execution inside existing systems, and finally maintenance — agents keeping a system running over time. Agentic use takes individual chat workflows and moves them into systems-level work that impacts more than the individual.
In chat, legal work is 57% writing and 20.5% knowledge retrieval. In agentic use, coding jumps to 32.9%, system operations to 17.7%, and workflow automation to 7.7% — agents comparing terms, flagging deviations, drafting red lines, and monitoring commitments, while humans keep negotiating material terms, setting risk tolerance, and owning accountability. The same pattern shows up in basically every department.
Even frontier users climbing into execution and maintenance are mostly running agents in individual silos. The next generation of gains will come at the intersection of different teams — team AI rather than just individual AI. And with frontier firms racing ahead, the numbers should be a wake-up call for anyone not deploying agents at scale yet.
Per roadmap documents viewed by The Information, Meta is finishing a consumer agent known internally as Hatch — a streamlined, OpenClaw-style agent experience that could anchor a new AI agent subscription justifying a $200-a-month price tag for high-usage accounts. The rollout could begin as soon as this week as a preview to a small group of customers.
Meta plans a new platform on WhatsApp for third-party agent integration, letting multiple agents coordinate with each other via WhatsApp messages. It mirrors some GrokBot functionality — but this isn't cribbing, these are just interaction patterns we're likely to see much more of.
Back in July, AI chief Alexandr Wang told staff that Watermelon had already caught up with GPT-55 on internal benchmarks. But the frontier has since moved with GPT-56 and will likely move again by October — so the question remains whether Meta gets closer to the frontier without actually reaching it.
GrokBot's launch pricing was intentionally rate-limiting — users initially weren't sure whether they needed both Cursor Ultra and SuperGrok Heavy, a $500-a-month combination. As of this week it's included in the $60-a-month Cursor Pro subscription and the $100-a-month SuperGrok subscription.
GPT-56 Sol drops to $4 per million input tokens and $20 per million output, from $5 and $30, following cuts to Luna and Terra last month. Many read it as pressure on Anthropic ahead of its IPO, but the more likely driver: business customers won't just use the most expensive state-of-the-art model for everything anymore, and to the extent OpenAI has the compute to deliver frontier models cheaper, it makes sense to do so.
The company is now generating more than $150 million in annualized revenue, up 50% from two months ago, even though 97% of users access the platform entirely for free. For acquirers this isn't a revenue-multiple conversation: as the backbone of open weights challenging frontier gatekeeping, some argue it should be worth three to four times the $13 billion asking price.
In a single week: a deal to license technology and acquire talent from Poolside, a stake in data-labeling company Mercur, and a potential investment in Perplexity at a $30 billion valuation — on top of NeoCloud stakes, land and power deals, and data center backstops. Some now conceptualize NVIDIA as the central bank of compute, standing behind the AI economy and setting the price of the key resource; comparisons range from John Malone's 1980s cable empire to Alphabet's Other Bets.
NVIDIA doesn't operate its own fabs, so growing chip revenue is limited by supplier constraints it can't control. Turning investment outward — $42.3 billion in private companies as of March, with a higher number certain this week — supports the entire AI economy and, in turn, keeps NVIDIA's revenues strong for years to come.
Taiwanese prosecutors indicted nine people over a scheme to smuggle Blackwell 300 systems into China, including a manager in NVIDIA's distribution business and two Supermicro employees. The group allegedly ordered 130 Supermicro servers, with 74 delivered to China and 56 stopped — under 10,000 chips total, not nothing, but not enough to build a frontier training cluster.
A voluntary internal spreadsheet obtained by Business Insider shows extremely jagged AI usage: Azure ranged from $1 to $7,500 monthly, Cloud and AI topped out at $15,000, and one person in Customer and Partner Solutions ran up $28,000. Medians clustered around $150–$500 (Core AI the outlier at $975) — and notably, BI found no correlation between token burn and compensation or promotions.