AI exploded in a year, and people could not keep up: Stanford AI Index 2026

Stanford HAI's 2026 AI Index finds that AI capability exploded within a year while institutions and labor could not keep that pace. Accuracy on Humanity's Last Exam rose from 8.8% to 38.3% in a single year, a 29.5 percentage point jump, and 2025 private AI investment reached 581 billion dollars, more than double the prior year's 253 billion. ASAP summarizes the result from the primary source.

The hardest test cracked open in one year

AI capability surged within a year even on the hardest tests. Humanity's Last Exam is the hardest current AI benchmark, a set of about 2,500 questions written by experts. Its accuracy rose from 8.8% in 2025 to 38.3% in early 2026, a 29.5 percentage point gain. Top models also crossed human baselines in more areas, from PhD-level science questions to competition mathematics.

How to read that 29.5 point jump

The raw number quadrupled, yet the absolute level still sits at 38.3%. In other words, models still get more than 60% of these expert-designed questions wrong. That is why the "surpassed humans" story and the "still failing" story hold true in the same table. What matters is not the remaining score but the slope. If a benchmark saturates at this pace, a year from now this exam will likely outlive its usefulness as a difficulty gauge. We have entered the stretch where the ruler of capability is being outrun by capability itself.

Money moved faster than capability

Private AI investment in 2025 reached 581 billion dollars, more than double the prior year. The figure jumped from 253 billion dollars in a single year, and the United States accounted for 344 billion of it. Capital is pouring into AI even faster than the capability curve.

The frontier is effectively private territory

More than 90% of the notable models in 2025 came from industry. Industry produced 87 notable models, while all other sources such as universities and governments produced just seven. Frontier model development has effectively concentrated in a handful of companies.

The more capability exploded, the more entry seats vanished

The labor market is moving the opposite way even as AI capability explodes, according to Stanford HAI. According to Stanford HAI, hiring of entry-level software developers fell, while mid-career and senior roles held steady or grew. Acceleration in capability did not translate directly into more jobs.

What these numbers mean for the Korean market

The fact that 344 billion dollars in investment clustered in the United States means that, for Korea, the race to match frontier models with domestic capital is effectively closed. The door that stays open is the application, data, and localization layer. The heavier signal is the drop in entry-level developer hiring. Korea's junior developer recruiting, bootcamps, and degree programs are still designed around an assumed baseline of entry-level coding skill, and that assumption is wobbling. What AI displaces is not code but the first rung of a career, and if that rung disappears, the pool of mid-career and senior talent eventually dries up too, a lagged risk that outlasts today's headlines.

What this report does not say

There is a caveat. Investment, model counts, and benchmarks are all things that are easy to measure. Harder to isolate from these figures alone is the real productivity contribution of deployed AI, the cost of its errors, and whether the fall in entry-level hiring is truly due to AI or to the macroeconomy. The definition of a "notable model" is itself a line drawn by the observer. So this report shows the direction of acceleration sharply and the causation of its results only faintly. The diagnosis that a gap exists is trustworthy, but pinning that gap to a single cause tips into over-reading.

The question that remains is a human one

The central question Stanford HAI poses in 2026 is whether the systems built around AI can keep up with its pace. Capability, investment, and concentration all steepened, but labor, institutions, and public trust did not move at the same speed. Human-centered AI calls not for faster models but for designs that let people and institutions absorb that speed. Closing the gap is the work of institutions and people, not of the models.

Source: Stanford HAI, "2026 AI Index Report" (2026; Humanity's Last Exam 8.8 to 38.3%, private investment 581 billion dollars, over 90% industry models). Cross-check: IEEE Spectrum, "Stanford's AI Index for 2026" (2026).