Stanford AI Index 2026: 20 Numbers That Matter

The Stanford AI Index 2026 is the year’s most cited reference on where artificial intelligence actually stands — a 400-plus-page audit, published April 13, 2026, that spans investment, technical performance, the economy, responsible AI, science, education, policy, and public opinion. Almost everyone cites it; almost nobody reads all of it.

So we did the reading and pulled out the 20 numbers that actually change how you should plan. The headline figure is hard to ignore: global corporate AI investment reached $581.7 billion in 2025, up roughly 130% in a single year — more than double the prior record. But the more useful story is in the contradictions: a US-China performance gap that has all but closed, agents that leapt from 12% to 66.3% on computer tasks, and frontier models that still read analog clocks correctly only about half the time.

This is the cross-domain digest. For the agent-specific deep dive, see our 200-data-point reference on the state of AI agents; for the coding-only numbers the Index surfaced, see our 50 verified AI coding statistics. Below, every figure carries its chapter provenance and a plain read on what it means for your business.

The AI Index is produced annually by the Stanford Institute for Human-Centered AI (HAI). The 2026 edition, released on April 13, is the canonical current reference: it aggregates data from Epoch AI, the International Federation of Robotics, benchmark trackers, government records, and original surveys into nine chapters. Its value is breadth and rigour — it is the closest thing the field has to a shared scoreboard.

That breadth is also why it is hard to use. Statistics get pulled out of context constantly: the $581.7 billion headline is corporate investment (private funding plus mergers, acquisitions, and minority stakes), not private investment alone, which is $344.7 billion. Population adoption of 53% is a global figure measured within three years of availability — the US specifically sits at just 28.3%. We flag those distinctions as we go, because the misreadings are where bad plans get made.

Here is the entire digest in a single scannable table, grouped into four themes — money, the frontier, the workplace, and the trust gap. Each row pairs the headline number with its year-over-year context and a one-line business implication. Read down the column that matches your decision; the surrounding sections unpack the figures that need the most care.

That is the report at a glance. The rest of this digest interprets the four numbers that are most often miscited or most consequential for a business: the true shape of the investment surge, the jagged frontier, the US-China convergence, and the widening gap between what experts and the public believe about AI.

The investment numbers are the cleanest signal in the report, and the most easily mangled. Global corporate AI investment reached $581.7 billion in 2025 — up about 130% from $253 billion in 2024, and past the previous high of $360 billion set in 2021. Of that, private investment was $344.7 billion (up 127.5%), now 60% of the total, with generative AI alone growing more than 200% and capturing nearly half of all private AI funding.

The geography is lopsided. US private AI investment of $285.9 billion was 23.1 times China’s $12.4 billion. But the report explicitly cautions that this understates China: government guidance funds are estimated to have deployed roughly $184 billion into AI firms between 2000 and 2023, outside the private-investment tally. Read the 23-to-1 ratio as a private-capital gap, not a total-spend gap.

One private round captures the scale of this surge better than any aggregate: the era’s defining funding events include rounds like Anthropic’s $65B Series H, the kind of single-company raise that would have been the entire sector’s annual total a decade ago. Our reading: the 130% year-over-year jump is not a forecast anyone had on their books, and it puts real near-term pressure on compute pricing. If your AI project ROI model assumes flat inference costs, rebuild it — the binding constraint right now is capacity, not capability.

The single most useful framing in the 2026 report is what researchers call jagged intelligence. In the same window that Gemini Deep Think scored a gold-medal result at the 2025 International Mathematical Olympiad, leading models still read analog clocks correctly only about 50% of the time on ClockBench — versus 90.1% for humans. Capability is not a single dial that turns up uniformly; it is a ragged edge that is superhuman in some places and below a child’s level in others.

The agent numbers show the same pattern from a different angle. Success on OSWorld computer tasks jumped from roughly 12% to 66.3% in a single year — within six points of human performance. On Terminal-Bench, real-world task success rose from 20% to 77.3%. Yet robots succeed at only 12% of real household tasks like folding clothes, despite scoring 89.4% on the same tasks in simulation. The lab-to-reality gap is still enormous.

The most strategically important convergence in the report is the US-China model performance gap, which has narrowed to 2.7% as of March 2026 — down from a gap of 17.5 to 31.6 percentage points in May 2023. US and Chinese models have traded the top position multiple times since early 2025; in February 2025, DeepSeek-R1 briefly matched the leading US model. The top US model now leads by less than three points.

On model output, the US still leads on the count of notable releases — the report’s figures put US releases in the range of roughly 50 to 60 for 2025, with the discrepancy coming from how Epoch AI defines a “notable” model across the report’s landing page and its R&D chapter. China released about 35. Industry produced more than 90% of all notable frontier models, up from under 50% a decade ago. The headline is not who is ahead, but how little daylight is left.

For procurement, the takeaway is concrete. When capability is within three points and several of the strongest models ship as open weights, the deciding factors shift to cost, latency, data sovereignty, and governance rather than raw benchmark supremacy. That is a different shortlist than the one most enterprises were running 18 months ago, and it is worth revisiting deliberately rather than defaulting to last year’s vendor.

Organizational AI adoption reached 88% of surveyed organizations in 2025, with 70% using generative AI in at least one business function. But agent deployment — the autonomous, multi-step kind — remains in single digits across nearly every business function. That gap between “we use AI” and “we have agents doing work” is the single clearest competitive opening in the whole report.

On productivity, the report is careful and so are we: studies report gains of roughly 14-15% in customer support, 26% in software development, and 50% in marketing output. These come from multiple independent studies with differing methodologies, not a single meta-analysis — treat them as directional, and note that the report also flags concerns about long-term learning penalties from heavy AI reliance and smaller gains on tasks that require deeper reasoning.

Our forward read: the adoption-versus-agent gap will be the defining competitive story of the next two years. Generic AI use is now table stakes — your competitors have it too. The advantage accrues to teams that move from chatting with a model to wiring agents into real workflows, carefully and on well-bounded tasks first. The labor data says the disruption is “targeted and just beginning,” concentrated in entry-level and structured roles, which means the organizations that win will be the ones that redesign those roles deliberately rather than simply cutting them. For coding-specific adoption numbers, our AI coding statistics digest breaks down the SWE-bench and developer-employment data in detail.

The report’s most actionable risk flag is the transparency paradox. The Foundation Model Transparency Index — a separate index whose results the Index cites — saw its average score drop to 40 points from 58 the year before. The most powerful models, from the leading labs, no longer disclose training code, dataset sizes, or parameter counts. Capability and openness are now moving in opposite directions.

Two more numbers compound the concern. Documented AI incidents rose to 362 in 2025, up from 233 in 2024, while reporting on responsible-AI benchmarks remains spotty even as nearly all developers report capability benchmarks. And there is a yawning perception gap: 73% of AI experts expect a positive job impact, versus just 23% of the general public — a 50-point chasm. Globally, 59% feel optimistic about AI’s benefits, but 52% say AI products make them nervous, and only 31% of Americans trust their government to regulate AI, the lowest of any country surveyed.

The expert-public perception gap is the one most business leaders underrate, and it is business-critical. A rollout that excites your engineering team can genuinely unsettle staff and customers who sit on the other side of that 50-point gap. Change management, not model selection, is where most AI initiatives actually succeed or fail — the data says the technology is ready well before the people around it feel that way.

A digest is only useful if it changes a decision. Here is how the four themes map to concrete moves for a marketing or product team planning the next two quarters.

The Index is a snapshot, not a strategy. Turning its numbers into a roadmap — a fresh model evaluation on your own corpus, a narrow agent pilot, and the governance scaffolding around both — is exactly the kind of work our AI and digital transformation engagements are built for, and our agentic SEO work applies the same discipline to search and content. The data is the easy part; acting on it deliberately is the differentiator.