Summary
- OpenAI's Enterprise Signals shows the top 10% of companies now generate 8.3 times more tokens per active user than typical companies, up from 2.6x in January
- Basis cut onboarding time from 2 hours to 30 minutes, while Clay used per-account subagents to save about an hour a day on nightly inbox cleanup
- Exa Labs lets Codex handle everything from spotting integration opportunities to writing PRs, running tests, and drafting weekly updates, while keeping human review before deployment
According to Enterprise Signals, a report OpenAI recently published, companies in the top 10% for AI usage — the "frontier" companies — now produce 8.3 times more output tokens per active user than typical companies. Back in January, that gap was only 2.6x. In half a year, it's nearly tripled. OpenAI frames this not as a matter of using AI more, but as a difference in how deeply companies wire agents into company context and tools, and how much real work they hand off.
Why the gap tripled
The metric OpenAI published sets April 2025 as the baseline index value of 1x, then tracks the output tokens generated per active user each month afterward. By June 2026, frontier companies had climbed to 17.1x and typical companies to 2.1x — divide one by the other and you get the 8.3x gap. OpenAI points to this not as simple growth in usage volume, but as a shift in how work gets organized: connecting agents to company context and tools, handing them heavier tasks, and turning workflows that succeed into repeatable processes.
OpenAI cites three startups as examples of this trend: Basis, which builds AI agents for accounting firms; Clay, which builds automation engines for sales and marketing teams; and Exa Labs, which builds web search infrastructure for AI agents. The industries differ, but the approach looks similar across all three — first teach the agent a stable procedure, keep feeding it context as the work changes, and eventually let it turn identified opportunities into verified execution.
Basis: turning onboarding into a teachable procedure
At Basis, first-day onboarding used to take two hours. Now it takes thirty minutes. On day one, a new employee gets immediate access to Codex and a company-specific onboarding skill. Codex handles the welcome, walks through core concepts, and processes setup work like account linking in the background. When recurring questions or edge cases come up, HR can revise the skill right away for the next cohort. Onboarding no longer depends on one person's schedule, but a human still steps in for exceptions.
Clay: giving scattered sales context a home
Clay is tackling a familiar problem — the context needed to close a deal is scattered across CRM, email, Slack, calls, presentations, texts, and internal team conversations. One of Clay's GTM engineers tried giving each account a persistent workspace and a dedicated subagent. Overnight, the subagent scans primary sources and updates the deal folder; in the morning, a coordinating agent turns those updates into a prioritized list. Clay reports this workflow cut about an hour a day off the time reps used to spend clearing their inbox each night, and because every recommendation comes attached to its source material, sales reps can check the reasoning themselves before acting on it.
Exa Labs: turning discovered opportunities into verified execution
Exa Labs, which builds a search API for AI agents, has set a goal it calls "Exa everywhere" — getting its API attached wherever developers happen to be working. Previously, developer relations and account teams had to manually comb through repositories and ecosystems to spot promising integration opportunities. Now Codex continuously monitors for high-priority integration opportunities, gathers relevant context, opens pull requests, runs tests, and prepares weekly updates drawn from Slack and Notion. When needed, it even drafts announcement copy for the team to review. Humans still decide which opportunities matter, what commitments to make, and how to manage external relationships.
What the three cases have in common
| Company | What the agent handles | Where context comes from | Where humans step in |
|---|---|---|---|
| Basis | New-hire onboarding guidance, account linking | Company-specific onboarding skill | Exception questions, skill revisions |
| Clay | Organizing deal context, surfacing priorities | CRM, email, Slack, call records | Verifying evidence before acting |
| Exa Labs | Finding integration opportunities, writing PRs, testing | Repositories, Slack, Notion | Review and approval before deployment |
All three companies treat exceptions as material for refining the next execution. Basis revises its skill based on recurring exception questions; Clay keeps context current through new account activity and rep verification; Exa Labs narrows the boundaries of future execution using test results and human review.
What companies can try now
OpenAI laid out six steps leaders can use as a reference for this shift: pick one task with a clear track record and high repeat frequency; set goals and measurement criteria for it (cadence, quality, cost, risk); define when the agent should start work and when it should stop and hand off for human review; clarify the roles of the people around the agent (business outcomes, access rights, adoption responsibility); repackage the experiment as a reusable skill, plugin, or shared workspace; and carry successful operating patterns over to the next task. OpenAI also shared its own research finding that, six months into adoption, junior employees sent 13 more messages per week to ChatGPT than executives did — underscoring that giving people room to experiment actually drives results.
This builds on a pattern OpenAI showed earlier in a tutorial titled ChatGPT Work Demo, From Meeting Briefs to Strategy Decks, which walked through attaching Work mode to sales, marketing, and strategy tasks.
Editor's take
This gap disclosure isn't just a status update. When OpenAI says frontier companies use 8.3 times more tokens, that also means their Codex and Work subscription bills are growing by roughly the same multiple. So this announcement lines up neatly with a product strategy that says: wire your agents deep into your work context. It's no coincidence that all three startups OpenAI cites as examples use Codex or ChatGPT as their core execution tool.
Anyone who watched the early days of ERP and CRM adoption will recognize this pattern. Early on, only a handful of departments used the system deeply while everyone else touched it superficially — and a few years later, that gap had turned into a gap in organizational competitiveness. Agent adoption appears to be following the same curve. The difference is speed: this time the gap tripled in just six months.
For companies here, the practical takeaway is this: don't treat raw token usage as a KPI. Measure results with concrete operational metrics instead — onboarding time, inbox cleanup time, number of PRs processed. As the Basis and Clay cases show, automation shouldn't target the whole company at once; start with one high-frequency task. And building a structure that turns exceptions into material for improving the skill, from the very start, will save you a lot of manual work down the line.
In the months ahead, OpenAI will likely use this kind of gap data to further segment Work and plugin pricing by company size, and expand Codex's continuous-monitoring capabilities into examples from other industries. The wider the gap grows, the stronger OpenAI's incentive becomes to keep producing consulting-style content aimed at pulling lagging companies up to speed.
Comments