Karpathy estimates only 5,000 people worldwide have access to internal frontier AI systems

Andrej Karpathy estimates that roughly 5,000 people worldwide have access to internal frontier-grade AI systems, a group he puts at about 0.00006 percent of the global population. In a post on X, summarized by The Decoder on Oct 12, 2026, he argues that public perception of AI capability is falling far behind what the most advanced systems can do. The OpenAI co-founder now works at Anthropic, according to The Decoder.

All user counts below are Karpathy's own estimates. He did not cite measured data, and the descriptions of internal systems are unverified.

Karpathy's four tiers of LLM exposure

1. About 6 billion people (~75% of the global population): Have barely come into contact with LLMs.
2. 1-2 billion casual users: Use free-tier ChatGPT-style products "a bit like a better Google search, a writing aid, or etc." Karpathy says this tier includes many professionals outside math and code. Executives, for example, mostly talk to other people, so their understanding is "second-hand and a bit abstract."
3. About 20 million professionals (~0.2%): Use frontier-grade LLMs for math and programming. Karpathy says they see "large, complex projects that used to take them weeks/months" completed by agents from a prompt. His examples are "building apps, copying apps, translating apps, decompiling apps from binaries."
4. About 5,000 people: Have access to internal frontier-grade systems.

What the top tier reportedly sees

Karpathy writes that "the external world has seen the preview." He describes the internal view as "swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics." He says this is work that would have taken top industry professionals years.

He also says "human review and comprehension are starting to fall behind." As evidence, he points to people still doing "archeology" on the OpenAI-HF incident from many months ago. He adds that mathematicians "may be poring over the 722 manuscripts on frontier mathematics for a while." The source article does not explain either reference, and neither claim has been independently confirmed.

Karpathy's own shift

The Decoder notes that in October 2024 Karpathy called AI agents mostly marketing that barely worked. Within months, he says, he stopped writing code by hand and let agents handle it.

Why coding and math lead

Karpathy says capability currently peaks in domains that are "digital, verifiable and economically valuable." He attributes this to capability that "emerges from reinforcement learning on verifiable rewards," and to companies putting training resources into the highest-revenue domains. Right now, that means coding and math.

Altman makes a similar claim

According to The Decoder, OpenAI CEO Sam Altman made a comparable argument in early September at the G20 Innovation Ministerial. Altman said a current AI coding tool could achieve in "17 minutes" what a startup had previously built over three months, and that "the world has not caught up with what this is going to mean." This is Altman's claim, and the startup and tool were not identified in the source.

What this means

The tier estimates are rough, but they match a pattern visible in adoption data: usage depth varies widely. The most consequential claim is the one that can't be checked. Neither Karpathy nor Altman has published evidence of what internal systems do, so outsiders can't verify a gap that two prominent figures from competing labs both describe.

The point about verifiable rewards matters more for practitioners. If capability tracks domains with checkable outputs and high revenue, progress will be uneven. Code, math and security will advance faster than fields without clean reward signals. That affects where builders should expect agent reliability to improve first.

Karpathy's warning that review is lagging is also a concrete problem. When agent swarms produce work faster than humans can audit it, evaluation, verification and oversight become the bottleneck. Expect that to shape product and policy discussion more than raw capability numbers.

No new model, pricing or benchmark data accompanied the post.