Stanford AI Index 2026: 12 Findings, Dissected

TL;DR

Stanford HAI’s 2026 AI Index is a 400-page snapshot of where AI stands right now. The headline numbers: coding benchmarks jumped from 60% to near-100% in twelve months, entry-level developer hiring fell almost 20% since 2024, global corporate AI investment hit $581 billion, and the gap between US and Chinese frontier models shrank to 2.7%. Meanwhile, only 31% of Americans trust their government to regulate any of it. The full report is free, but if you’d rather spend 10 minutes instead of 10 hours, this is the version that strips the padding and adds the context Stanford leaves out.

What the Stanford AI Index Actually Is

The AI Index is an annual report from Stanford’s Institute for Human-Centered Artificial Intelligence (HAI). The 2026 edition (the ninth) tracks hundreds of metrics across AI performance, investment, adoption, talent, regulation, and public opinion. It pulls from academic databases, patent offices, job boards, government filings, and company disclosures.

I’ve read every edition since the first one in 2017. The early reports felt academic: careful charts, measured conclusions, lots of caveats. This year’s edition reads differently. The data is moving so fast that even Stanford’s cautious framing can’t hide the tectonic shifts underneath. Coding agents going from “interesting demo” to “near-human” in a single calendar year, and a 20% contraction in entry-level developer positions. Both are inflection points, and the report treats them accordingly.

Here are the 12 key findings, grouped by what they actually mean for people who build software.

AI Performance: From 60% to Near-100% in Twelve Months

The most jarring chart in the entire report sits on page 47. On SWE-bench Verified, a benchmark where models resolve real GitHub issues, the best score was 60% in early 2025. By the end of the year, it was close to 100%. No benchmark in AI history has moved that fast.

On Humanity’s Last Exam (HLE), a test designed specifically to stump frontier models, top scores went from 8.8% in 2025 to over 40% by April 2026. Claude Opus 4.6 scored 34.4% on the official leaderboard, and Gemini 3.1 Pro hit 46.4%. Google’s Gemini Deep Think won a gold medal at the International Mathematical Olympiad.

AI agents on OSWorld, a benchmark where models interact with real desktop applications, improved from 12% to roughly 66% task success over the same window. The curve is unmistakable: we crossed the threshold where AI systems can do real work beyond toy benchmarks.

If you want to verify the SWE-bench numbers yourself, the dataset is public. Here’s how to pull the verified subset and check a model’s score:

from datasets import load_dataset

swe_bench = load_dataset("princeton-nlp/SWE-bench_Verified", split="test")
print(f"Total verified instances: {len(swe_bench)}")
print(f"Repos covered: {len(set(r['repo'] for r in swe_bench))}")
print(f"Sample task: {swe_bench[0]['instance_id']}")
# Total verified instances: 500
# Repos covered: 12
# Sample task: astropy__astropy-12907

The benchmark contains 500 real GitHub issues across 12 popular Python repos. When the report says “near 100%,” it means frontier models can now patch nearly all of them autonomously, including file localization, code generation, and test validation.

And then there’s the humbling counterpoint. The best model tested, GPT-5.4, reads analog clocks correctly 50.1% of the time. Claude Opus 4.6 scored 8.9%. These systems can solve differential equations but struggle with something every second-grader can do. The report doesn’t dwell on this, but it’s a useful reminder: these models are not generally intelligent. They’re pattern-completion engines with specific blind spots that would be comical if people weren’t betting their infrastructure budgets on them.

The US-China Race: A 2.7% Lead on a 23:1 Budget

The report confirms what people in the industry have been whispering for months. The performance gap between American and Chinese frontier models has collapsed. As of March 2026, Anthropic’s top model leads the best Chinese model by just 2.7% across a basket of benchmarks.

The investment gap has not collapsed. The US poured $285.9 billion in private AI investment in 2025. China spent $12.4 billion. That’s a 23:1 ratio. China is producing competitive models at a fraction of the cost. The report doesn’t explain the mechanism.

China leads one category decisively: industrial robots. Beijing installed 295,000 in 2024 versus America’s 34,200. China is also outpacing the US on research publication volume, though publication count and research quality are different things.

The talent pipeline tells a worrying story. The number of AI researchers moving to the US dropped 89% since 2017, with an 80% decline in just the last year alone. If this trend continues, the US investment advantage becomes a pile of GPUs with fewer and fewer people who know how to make them productive.

$581 Billion: Where the Money Went

Global corporate AI investment hit $581 billion in 2025, smashing the previous record of $360 billion set in 2021 and more than doubling 2024’s $253 billion. The US accounted for $344 billion of that total, nearly 60%.

The concentration is stark — we covered this dynamic in our AI infrastructure and layoffs analysis. Five hyperscalers now control more than two-thirds of global AI compute. Nvidia alone holds over 60% of global AI compute capacity, with Amazon and Google in second and third place. World AI compute capacity has expanded 3.3x annually since 2022, growing 30-fold since 2021.

GitHub tracked 5.58 million AI projects in 2025, a 23.7% jump from 2024 and roughly 5x the number from 2020. Stanford counted 1,953 newly funded US AI companies in 2025.

On the consumer side, the report estimates generative AI delivers $172 billion in annual value to US consumers as of early 2026, with the median value per user tripling from 2025 to 2026.

The Entry-Level Developer Question

This is the finding that landed in every LinkedIn feed. Entry-level software developer positions for workers aged 22-25 fell nearly 20% since 2024. The Stanford report calls it “the first white-collar job category to show measurable contraction attributable to AI.”

The productivity numbers explain the mechanism. The report aggregates multiple controlled experiments that found 14-26% productivity gains in software development and customer support, and up to 72% in marketing. A team of five senior engineers with AI tools can produce what previously required eight people, including three juniors.

But the report also contains a finding that complicates the “AI is killing jobs” narrative: workers in occupations least exposed to AI experienced greater unemployment increases than those most exposed. The people using AI the most are not the ones losing their jobs. The people in roles AI can’t touch, the ones being restructured for other reasons, are.

I wrote about the junior developer hiring contraction last month using BLS and Indeed data. The Stanford report adds the authoritative stamp. The question is no longer whether it’s happening but what replaces the traditional junior-to-senior career pipeline when companies stop hiring at the bottom rung. Our SWE job market analysis digs into the two-tier split in more detail.

Meanwhile, the AI-adjacent job market is overheating. Agentic AI and AI agent postings grew roughly 280% year over year. AI governance roles grew 17%. The PwC Global AI Jobs Barometer puts the wage premium for AI-skilled workers at 56%, up from 25% the prior year. The 2026 Dice report shows AI skills appearing in 73% of US tech job postings, up 192% year over year.

The market is bifurcating. AI-fluent workers face record demand while roles that don’t touch AI are thinning out fast.

Adoption: Faster Than the PC, Faster Than the Internet

Generative AI hit 53% population-level adoption in three years. For context: the personal computer took about a decade to reach that level. The internet took roughly seven years. Only the smartphone, at about five years, came close.

Organizational adoption reached 88% in 2025. Separate industry surveys (G2, CrewAI) put enterprise AI agent adoption at 57-65%, with most deployments still in pilot rather than full production.

The adoption curve is steepest in education. Four out of five US students use AI for schoolwork. Only 50% of schools have any AI policy in place, and only 6% of teachers say those policies are clear. The gap between how students use AI and how institutions manage it is enormous.

Over 90% of notable frontier models released in 2025 came from industry, not academia — up from roughly half in 2015. The report tracks 87 out of 94 notable 2025 models as industry-produced. Universities are increasingly consumers of AI rather than producers of it.

Transparency, Trust, and the Governance Gap

Model transparency is declining. The Foundation Model Transparency Index fell from 58 to 40 points in the past year. As models get more capable, the companies building them share less about training data, methodology, and safety evaluations. Most AI companies don’t publish detailed bias audits.

Public trust tells its own story. Seventy-three percent of AI researchers believe AI’s impact on labor will be positive. Only 23% of the general public agrees. That 50-point trust gap between experts and the public is the widest the Index has ever recorded.

The US ranks dead last among surveyed nations in public trust that government can regulate AI: 31%. Singapore leads at 81%. Southeast Asian nations show the most positive sentiment overall, while Germany, France, and the Netherlands showed the strongest year-over-year improvements in public perception, gaining 10-12 percentage points each.

Documented AI incidents hit 362 in 2025, up from 233 in 2024 (a 55% increase). The report doesn’t editorialize much here, but more AI in the wild means more things going wrong, and the regulatory frameworks aren’t keeping pace.

The Environmental Bill

Training Grok 4 produced an estimated 72,816 tons of CO₂ equivalent, with some alternative estimates reaching 140,000 tons. For comparison, GPT-4’s training was estimated at 5,184 tons. Llama 3.1 405B clocked in at 8,930 tons. The carbon cost of frontier model training is growing faster than the models’ capability improvements.

AI data centers worldwide draw 29.6 gigawatts of power, equivalent to New York City’s peak electricity demand. The most efficient inference models consume about 5 watts per response; the least efficient exceed 50 watts. That gap is worth paying attention to because inference (running the models) dwarfs training in total energy consumption once a model ships to millions of users.

The report doesn’t make policy recommendations. But the implication is hard to miss: the current AI buildout is a massive energy bet. If the productivity gains justify the power draw, it’s a worthwhile trade. If they don’t (and the jury is still out for most use cases), the environmental cost is baked in whether the returns materialize or not.

What the Report Misses

Stanford’s Index is data-rich but deliberately avoids strong conclusions. A few gaps the report undercounts or omits:

The report tracks notable model releases but doesn’t weight the downstream effect of open-weight models like DeepSeek, Llama, and GLM. China’s ability to close the performance gap at 1/23rd the investment likely relies heavily on open-weight foundations that circulate freely regardless of export controls. We covered the economics of this in our GLM-5.2 review.

On regulation, the report covers US and Chinese AI policy but spends less time on the EU AI Act, which reaches its high-risk system compliance deadline on August 2, 2026. For companies operating across both markets, the regulatory divergence is the bigger story.

The SWE-bench numbers are impressive, but the report doesn’t discuss how much it costs to run these coding agents at scale. Our FastContext analysis showed that coding agents waste roughly half their tokens on repo search — a cost that doesn’t show up in benchmark scores but dominates real-world inference bills.

FAQ

What is the Stanford AI Index?

An annual report from Stanford’s Institute for Human-Centered Artificial Intelligence (HAI) that tracks AI progress across benchmarks, investment, adoption, talent, policy, and public opinion. The 2026 edition is the ninth, running over 400 pages with data through early 2026.

How did AI coding benchmarks change in 2026?

SWE-bench Verified (where models resolve real GitHub issues) went from 60% to near 100% in a single year. AI agents on OSWorld improved from 12% to 66% task success. These are the largest single-year improvements in any AI benchmark category.

Are entry-level developer jobs declining because of AI?

Yes. The report shows entry-level software developer positions for ages 22-25 fell nearly 20% since 2024, making it the first white-collar job category with measurable AI-driven contraction. The cause: senior developers with AI tools produce enough output that companies need fewer juniors.

How much was invested in AI in 2025?

Global corporate AI investment hit $581 billion in 2025, more than doubling 2024’s $253 billion. The US accounted for $344 billion. Private AI investment specifically reached $285.9 billion in the US versus $12.4 billion in China.

Is the US still leading China in AI?

On benchmarks, barely: a 2.7% lead as of March 2026. On investment, massively: a 23:1 ratio. On talent pipeline, the trend is negative: AI researchers moving to the US dropped 89% since 2017. China leads on industrial robot deployment and research publication volume.

Sources

• The 2026 AI Index Report — Stanford HAI — the full 400-page report, primary source for all findings in this article
• Inside the AI Index: 12 Takeaways from the 2026 Report — Stanford HAI — Stanford’s own summary of the key findings
• Stanford’s AI Index for 2026 Shows the State of AI — IEEE Spectrum — IEEE’s independent analysis of the report’s data
• Stanford AI Index 2026: Developer Employment Down Nearly 20% — Tech Jacks Solutions — coverage of the entry-level hiring contraction finding
• Stanford’s 2026 AI Index: Where AI Actually Stands — Stark Insider — additional data analysis of the full report

Bottom Line

The Stanford AI Index is the closest thing the field has to an annual physical. This year’s checkup says the patient is stronger than ever and growing faster than anyone predicted. It’s also burning more energy than ever. Power is concentrating in fewer hands, transparency is declining, and institutions can’t keep up.

For developers, the numbers speak plainly. AI coding tools went from 60% to near-100% on real-world tasks in twelve months. Entry-level hiring dropped 20%. AI skill premiums hit 56%. All trailing indicators from data that’s already six months old — these numbers describe where you already stand.