OpenAI's capability overhang report reframes AI adoption as a usage gap, not an access gap

The metric that carries the argument: 'thinking capabilities' per person

The report's central claim rests on a specific, measurable unit: the amount of 'advanced thinking capabilities' a person uses — meaning AI applied to complex, multi-step work rather than one-off prompts. OpenAI reports that a typical power user relies on about seven times more of these capabilities than a typical user.

Osborne then extends the same measure across borders. Among more than 70 countries with the highest ChatGPT usage, some use three times more thinking capabilities per person than others. That is the 'overhang' — not a shortage of accounts or model access, but a gap in how deeply the tools are actually put to work.

This is a deliberate reframing. Adoption debates usually center on availability and cost; OpenAI is arguing the constraint has moved downstream to depth of use. The distinction matters for anyone building automation, because it separates 'we have the model' from 'we route real work through it.'

The country data cuts against the income story

The most concrete finding is that the gap 'isn't driven by income alone.' The United States and India lead in total users, while smaller high-income countries like Singapore and the Netherlands stand out for population penetration. But the agentic-usage leaderboard looks different.

Osborne names Vietnam and Pakistan as among the world's top users of agentic tools, with more than 2x higher per-person use of advanced tasks like data analysis, Connectors, and Codex. That is a pointed example: two lower-income countries outpacing wealthier ones on exactly the automation-heavy usage OpenAI is measuring.

The implication OpenAI draws is that behavior and fluency, not GDP, determine who captures productivity gains. Whether the per-person metric holds up across small user bases is worth scrutiny — but the direction of the claim is that adoption depth is learnable, which is precisely what makes it a policy target.

The response is institutional, and it starts with schools

OpenAI for Countries, launched last year, is being expanded in 2026 with initiatives across education, health, AI skills training and certifications, disaster response, cybersecurity, and start-up accelerators. The framing is a menu rather than a single product — partnerships 'shaped through ongoing discussions' rather than a fixed offering.

The lead initiative is Education for Countries, with a first cohort of Estonia, the United Arab Emirates, Greece, Jordan, Slovakia, Kazakhstan, Trinidad & Tobago, and Italy's CRUI. The program combines expanded tool access, research into AI's effect on learning, and certifications for both students and educators, working through ministries and universities.

With more workplaces adopting AI and more employers seeking workers with AI skills, governments are increasingly treating the technology as essential education infrastructure.Montana Labs

One line is worth flagging for what it reveals about incentives: the education program is designed 'to work hand-in-hand with governments to improve our models and education tools.' The partnerships are a channel for both distribution and data.

What this announcement actually sets in motion

Read closely, this is OpenAI positioning national governments — not individual enterprises — as the unit that closes the automation gap. By measuring the divide in agentic usage and then offering ministries a structured way to raise it, the company is tying its commercial growth to a public-policy narrative about who falls behind.

For teams building applied AI, the useful takeaway is the metric, not the mission. If a country-scale program treats depth of agentic use as the thing to optimize, the same logic applies inside any organization: the value gap is between people who chain AI across multi-step work and those who stop at simple prompts. Closing that gap is an adoption and fluency problem, and OpenAI has now put a number on how large it can get.