What happened
OpenAI released a report called Enterprise Signals, built from aggregated, de-identified usage data across its enterprise customer base. The report ranks companies each month by output tokens per active user, calling the top 10% "frontier firms" and the companies sitting between the 45th and 55th percentiles "typical firms." The headline number: as of June 2026, frontier firms generated 8.3 times as many output tokens per active user as typical firms, up from a 2.6 times gap in January. The gap is not evenly spread. It is widest in information and technology, at 11.7 times, and narrowest in manufacturing, at 5.3 times. Token growth among typical firms was modest across every industry OpenAI measured, between 1.9 and 2.8 times over the same period. A second finding sits underneath the headline gap. As of June, Codex, OpenAI's agentic coding and task-execution product, generated 64% of combined Codex and ChatGPT output tokens among enterprise customers. OpenAI frames this as a shift from assistance, asking a model questions, to delegation, handing a model a task and letting it work through multiple steps with access to tools and files. The fastest growth in Codex users since February came from outside engineering: legal use grew 108 times, sales and recruiting each grew 41 times, marketing grew 26 times, while engineering itself grew 5 times. Frontier firms also use context-connecting features at far higher rates. Plugins, which connect AI agents to a company's specific tools and data, are used weekly by 21% of active users at frontier firms versus 9% at typical firms. Skills, which let a company save and reuse task instructions, are used by 19% versus 3%.
Why it matters for business owners
This report is going to circulate. It comes directly from OpenAI, it landed the same week Anthropic was reported to be preparing a blockbuster public offering, and it gives every AI vendor, consultant, and conference speaker a clean, tripling statistic to put in a slide. A business owner who has been quietly using a chat assistant for drafting emails and summarizing documents is going to hear a version of this story that sounds like: everyone else is pulling ahead, and you are falling behind. That reaction is exactly the risk. A statistic built from token counts is easy to repeat and easy to misread as a call to spend more, buy more seats, or turn on more AI tools immediately. OpenAI's own report, read past the headline, is making a narrower and more useful point: the gap is not really about how much AI a company buys. It is about how deeply AI is actually connected to real work inside the business.
What owners should not misunderstand
This is not proof that using more AI tokens makes a company more productive. OpenAI's own report says so directly: "tokens are an imperfect measure of business value; a short response can be highly valuable, while a long one may add little." Ronnie Chatterji, OpenAI's chief economist, told Semafor that "high token usage doesn't necessarily equate to productivity," and that measuring the actual return on AI spend remains "frankly difficult to do." The company publishing the statistic is also the company telling you not to treat it as a scoreboard. This is not evidence that a small or midsize business is automatically behind because its token counts are lower than a large technology company's. The widest gap in the data sits inside the information and technology sector specifically, a sector defined by software work that is unusually well suited to agentic coding tools like Codex. A retail, real estate, or professional services business will not generate technology-sector token volumes by the nature of the work, and that difference does not mean the business is failing to adopt AI well. This is also not a report about which company bought the most AI licenses. Frontier firms are not distinguished by spend. They are distinguished by integration: connecting agents to company-specific tools and data through Plugins, and saving reusable task instructions through Skills, at roughly twice and six times the rate of typical firms. The gap tracks how AI is wired into the business, not how much of it was purchased.
The operational lesson
Vendor-published benchmarks like this one are useful for one thing and dangerous for another. They are useful for showing where the technology and the market are heading: toward agents doing multi-step work with real access to company tools and data, spreading well beyond software engineering into legal, sales, and marketing functions. They are dangerous when a business owner reads the topline multiple as a target to hit, rather than as a description of what leading firms actually did differently. What leading firms did differently, according to OpenAI's own breakdown, was give AI systems more context about the business and more permission to act on it, not simply run more prompts. A company that doubles its token spend on general-purpose chat without connecting any tool, file, or workflow to it will not close this gap. It will just spend more money asking the same kinds of questions.
What a serious business should do next
Before reacting to any AI usage benchmark, including this one, separate the volume question from the integration question. Ask how much AI your business actually uses only after you have asked a more specific question: which real, recurring tasks has your team actually handed off to an AI tool with access to the files, systems, or data needed to do that task well, rather than asked about in a general chat window. Audit what your team already has connected. If nobody in the business has set up anything like a Plugin, a Skill, or a comparable integration that lets an AI tool work with your actual data and tools, that is the real gap this report describes, not a shortfall in monthly token count. Closing it starts with one or two well-scoped workflows done properly, not a company-wide push to use AI more. When a vendor, consultant, or conference cites this kind of statistic to create urgency, ask what it is actually measuring before acting on it. A number built from token volume is a proxy for engagement, not a verdict on whether your AI spend, however large or small, is producing real value for your business.
The Atlacis view
A report like this is designed to travel fast and to create pressure. The headline multiple is real and it is verified. What gets lost is that OpenAI's own economist is telling anyone who reads past the first paragraph that the number is not a productivity measurement, and that proving AI spend is worth it remains genuinely hard to do even for OpenAI's own customers. Atlacis helps owners slow down before reacting to a statistic like this one. That means mapping which specific workflows in your business would actually benefit from an AI tool with real access to your data and tools, choosing a small number of them to do properly, and measuring whether the result changed the work, not whether the token count went up. Chasing a usage gap is a way to spend money. Closing a workflow gap, one task at a time, is how the spend turns into something a business owner can actually point to.
The short version
- OpenAI's new Enterprise Signals report found that frontier firms, the top 10% of monthly AI users, generate 8.3 times more output tokens per active user than typical firms, up from a 2.6 times gap in January 2026.
- A second finding: as of June 2026, Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers, a sign of AI shifting from answering questions to executing multi-step tasks.
- OpenAI's own report and its chief economist both caution that token volume is an imperfect measure of business value and that measuring real AI ROI remains difficult, even for OpenAI's own customers.
- What actually separates frontier firms is integration, connecting AI agents to company-specific tools and data through features like Plugins and Skills, not raw spend or token count.
- A business is not behind because its token counts are lower than a technology company's. The widest gap in the data sits inside the information and technology sector specifically.
- Audit which real, recurring workflows are actually connected to an AI tool with the access needed to do the task well, and close that gap one workflow at a time instead of reacting to a usage benchmark.
Where ATLACIS can help
Sources
- OpenAI: Enterprise Signals, what frontier firms are doing differently (updated August 12, 2026)
- Semafor: The gap is widening between corporate AI adopters and laggards (August 12, 2026)
- VKTR: OpenAI: Frontier Firms Use 8.3x More AI Than Typical Enterprises, Gap Tripling Since January (Michelle Hawley, August 12, 2026)