AI to ROI Reports & Data: The 2026 Stanford HAI 2026 AI Index Report
The most complete report on the State of AI you will read all year.
Stanford HAI 2026 AI Index Report
About the Authors
The Stanford Institute for Human-Centered Artificial Intelligence (HAI) has published the Artificial Intelligence Index Report annually since 2017. Its mission is to provide neutral, rigorously vetted global data on AI so that policymakers, executives, researchers, and the public can make more informed decisions. The 2026 edition is the ninth in the series and runs more than 400 pages across nine chapters.
The report is produced by an AI Index Steering Committee spanning academia, industry, and policy. Committee members include:
Stanford professors across bioengineering, economics, and computer science
The co-founder of Anthropic
Senior figures from Google, the Brookings Institution, and the University of New South Wales.
No single sector sets the agenda, which is why governments, media organizations, and major companies worldwide cite the report as a primary reference on AI.
Top 10 Findings in the Artificial Intelligence Index Report
1️⃣ AI Capability Is Accelerating, Not Plateauing
Frontier AI models are not slowing down:
Industry produced more than 90% of notable frontier models in 2025, and several now meet or exceed human baselines on PhD-level science questions, multimodal reasoning, and competition-level mathematics.
On SWE-bench Verified, the key coding benchmark, model performance rose from 60% to near 100% of the human baseline in a single year.
Organizational adoption reached 88% globally.
Four in five university students now use generative AI.
2️⃣ The US-China Performance Gap Has Effectively Closed
For years, US models held a commanding lead in performance rankings. That advantage is gone:
US and Chinese models have traded places at the top of the leaderboard multiple times since early 2025.
In February 2025, DeepSeek-R1 briefly matched the top US model. As of March 2026, Anthropic’s top model leads by just 2.7 percentage points.
The US still produces more high-impact patents and top-tier models overall.
China leads in publication volume, patent output, and industrial robot installations.
South Korea ranks first in AI patents per capita.
3️⃣ Investment Is Surging, but It’s Concentrated
Global corporate AI investment more than doubled in 2025, reaching $581.7 billion:
Private investment grew 127.5%, hitting $344.7 billion, with generative AI companies capturing $170.9 billion of that total.
The United States led all countries with $285.9 billion in private AI investment, 23.1 times the amount invested in China ($12.4 billion).
Within the US, California alone accounted for $218 billion, more than 75% of the national total.
The capital is real, but its distribution is narrow.
4️⃣ The Jagged Frontier: Brilliant and Blind in the Same Model
AI capability is not uniformly distributed, and the gaps are striking:
Gemini Deep Think earned a gold medal at the International Mathematical Olympiad, yet the best-performing model reads an analog clock correctly just 50.1% of the time.
AI agents on OSWorld, which tests performance on real computer tasks across operating systems, improved from roughly 12% accuracy to 66.3% in a single year, approaching human performance.
Meanwhile, robots succeed in only 12% of real household tasks, such as folding laundry or washing dishes.
5️⃣ Responsible AI Is Falling Behind in Capability
The infrastructure for responsible AI is growing, but it is not keeping pace with deployment:
Nearly all leading frontier model developers report results on capability benchmarks, but reporting on responsible AI benchmarks remains sparse.
Documented AI incidents recorded by the AI Incident Database rose to 362 in 2025, up from 233 in 2024.
The Foundation Model Transparency Index found that average disclosure scores dropped from 58 to 40, meaning the most powerful models are also the least transparent about how they were built.
6️⃣US Leads in Investment, but Lags in Attracting Global Talent
The United States remains the dominant destination for AI activity, with 5,427 data centers, more than 10 times that of any other country. But its ability to pull in global talent is collapsing:
The number of AI researchers and developers moving to the US has dropped 89% since 2017, with 80% of that decline occurring in the last year alone.
The majority of AI-related graduate students at US universities are non-residents, and federal policy changes around student visas are expected to push that number lower.
7️⃣ AI Adoption Is Growing, but the Value Is Unevenly Distributed
Generative AI reached 53% population adoption within three years of its mass-market introduction, faster than the personal computer or the internet.
US consumer surplus from generative AI reached an estimated $172 billion annually by early 2026, with the median value per user tripling between 2025 and 2026.
Yet adoption correlates strongly with GDP per capita.
The US ranks 24th globally, with adoption at 28.3%, while Singapore ranks 1st at 61%, and the UAE ranks 2nd at 54%.
8️⃣ AI Is Displacing Entry-Level Workers First
Productivity gains from AI are measurable in structured work:
14%-15% in customer support, 26% in software development, and 73% in marketing output. But those same fields are already seeing early-career employment decline.
US software developers aged 22 to 25 saw their employment fall by nearly 20% in 2024, even as headcount among older developers continued to grow.
Firm surveys indicate executives expect planned headcount reductions to accelerate.
The disruption is targeted and just beginning.
9️⃣ AI’s Environmental Cost Is Scaling with Its Capabilities
Training and running frontier models carry a significant and growing environmental footprint:
Grok 4’s estimated training emissions totaled 72,816 tons of CO2 equivalent, roughly equivalent to driving 17,000 cars for a year.
AI data center power capacity rose to 29.6 GW by the end of 2025, comparable to peak demand for all of New York state.
Annual inference water use for GPT-4o alone may exceed the drinking water needs of 12 million people.
The infrastructure underlying AI now rivals Switzerland’s national electricity consumption.
🔟 Expert and Public Views on AI Are Sharply Divided
AI experts and the general public hold fundamentally different views of the technology’s trajectory, and the gap is wide:
When asked about AI’s impact on how people do their jobs, 73% of AI experts expect a positive effect, compared to just 23% of the US public, a 50-point gap.
Similar divides appear in the economy and medical care.
Global optimism is rising overall, with 59% of respondents saying AI’s benefits outweigh its drawbacks, up from 52%.
But nervousness is rising too, up 2 points to 52%.
Only 31% of US residents trust the government’s ability to regulate AI.
What’s Next
The report points to a looming governance stress test. Organizational AI adoption is at 88%, and AI agents have begun handling real-world tasks. At the same time, the gap between what AI can do and what organizations can measure and oversee is widening:
Documented incidents are rising, transparency scores are falling, and companies still report knowledge gaps and regulatory uncertainty as the top two barriers to responsible AI implementation.
Boards and executive teams that treat governance as a compliance exercise will find themselves behind the firms that build it as an operating discipline.
AI sovereignty is becoming a defining element of national competition:
More than half of the national AI strategies adopted since 2024 came from emerging economies, and state-backed supercomputing investment is rising across every region.
At the same time, open-source development is redistributing participation, with contributions from outside the US and Europe now approaching US levels on GitHub.
Companies with global supply chains and multi-market operations should expect a patchwork of national AI requirements to become a standard feature of the regulatory environment by 2027.
Software development and customer support are the leading use cases for AI in the workplace now, but the report suggests it will broaden. AI agent deployment remains in single digits across nearly every business function, meaning the productivity effects visible so far are early and not very measurable.
As agent adoption scales, expect companies to invest in AI to augment experienced workers, while the number of entry-level roles is likely to decline in many industries.
Organizations that begin workforce planning now, including identifying which roles will be redefined rather than eliminated, will have more options than those that wait for the labor market to force the question.
The Report Summary Page, The Full 423-Page Report
Charts & Data for the Week
Software Pricing Models Are Changing Fast
The era of seat-based software pricing is on its last legs; however, picking the right pricing model, the one most aligned to customer value, isn’t an easy decision.
According to McKinsey, software companies have begun the shift to usage-based pricing by embracing activity-based metrics. The best activity-based metrics are linked to the business value generated. For example, McKinsey compares two AI SDR agents:
One monetizes based on effort expended (per thousand emails sent), while another charges per sales development workflow completed (leads identified).
The latter is closer to the business value that matters to the sales organization.
True outcome-based pricing takes things a step further, adding a success layer for each qualified lead.
McKinsey notes that, while outcome-based pricing may ultimately be the way to go, it is often challenging to execute. Making it work at scale requires several key factors:
The outcome being monetized must be highly correlated with the work the product does.
The lag time between workflow execution and the outcome should be short.
Customers should value the outcome roughly equally.
It must be possible to automatically confirm success without human assistance or judgment.
AI Is Deeply Integrated Into Our Workplaces Now
Organizational AI adoption rates are increasing rapidly, though at a slower pace than in the first half of 2025. 41% of employees say their organization has integrated AI technology or tools to improve organizational practices, up 3 points from the previous quarter.
Between Feb. 4-19, 2026, Gallup surveyed 23,717 U.S. employees about the state of AI work. The findings suggest that the growing presence of AI is reshaping workplace dynamics. Employees in organizations that have adopted AI are more likely to report disruption and both positive and negative changes in staffing levels. Employees who use AI frequently say it improves their productivity, but evidence that AI has fundamentally changed how work gets done across organizations remains more limited.
OpenAI Is Banking on Ads to Reach Its Revenue Goals
Despite its pivot to enterprise customers, OpenAI is going all in on advertising to meet its revenue projections through 2030. According to The Information, OpenAI expects advertising revenue to reach about $2.4 billion this year and quadruple to nearly $11 billion next year. A year ago, OpenAI’s ad revenue forecast was a more modest, but still very large, $1.6 billion in 2026 and $5.9 billion in 2027 from users on its free tier.
OpenAI believes that its ad engine will go into overdrive late in the decade. In 2030, OpenAI expects ads to generate about $102 billion, or 36% of its total revenue for that year. A year ago, it forecast that nonpaying users would generate only $26.5 billion in sales.
One Must-Read Story This Week - Misanthropic: On Mythos, Bad Human Behaviors and Systems Vulnerabilities by Michael Cembalest, J.P. Morgan Asset Management
About the Publication and Author
Michael Cembalest is Chairman of Market and Investment Strategy at J.P. Morgan Asset Management, where he writes the “Eye on the Market” series. Published April 13, 2026, this note was written over a weekend and shared internally before being distributed to clients. It draws entirely on publicly available information, with Cembalest noting that economic and market implications are still being worked through.
Article Summary
Mythos Performance
Anthropic’s new model, Mythos, surpasses all prior Claude models on key benchmarks. It achieves higher accuracy and token efficiency in web research tasks, outperforms predecessors on complex, long-horizon software engineering, and modestly accelerates Anthropic’s score on the Epoch Capabilities Index, a composite of 40 AI benchmarks. Mythos also scores best among Claude models on factuality across business, health, law, and STEM domains.
Anthropic describes Mythos as its best-aligned model yet, while simultaneously acknowledging it poses the greatest alignment-related risk of any model it has released. Cembalest notes this is a notable contradiction.
Cyber Vulnerability Detection and Project Glasswing
Mythos achieved a perfect score on Anthropic’s CyBench cybersecurity benchmark and scored 83% on CyberGym, versus 67% for Opus 4.6. When tested on Firefox, Mythos had a 72% shell exploitation success rate, compared to 1% for Opus 4.6. Its cyberhacking abilities are emergent, not by design. Mythos has identified thousands of high-severity vulnerabilities, including:
A 27-year-old flaw in OpenBSD allowing full network takeover without credentials
A flaw in the FFmpeg video encoder that survived 5 million automated tests
Vulnerabilities in the Linux kernel enabling complete machine takeover
In a 32-step corporate network attack simulation run by the UK AI Security Institute, Mythos was the first model to complete all 32 steps on its best attempt, averaging 22 steps, versus 16 for Opus 4.6 and 14 for GPT-5.4.
Anthropic will not release Mythos publicly, citing concerns about adversarial state actors. Instead, it launched Project Glasswing, granting 12 partner organizations (including Amazon, Apple, Cisco, Google, J.P. Morgan, Microsoft, and NVIDIA) access to find vulnerabilities in their own code and vendor software. Mythos Preview is priced at $25/$125 per million input/output tokens. Anthropic is providing $100 million in usage credits to Glasswing members.
A separate concern:
Operational technology systems in industrial settings often run 10 to 18 years without replacement, making them far harder to patch than cloud-based IT. J.P. Morgan estimates 20-30% of industrial network infrastructure has no available patch.
Disturbing Behaviors Found in Mythos
Anthropic’s behavioral audits show that Mythos cooperates with misuse attempts less than half as often as prior models did. Sycophancy and deception scores improved. However, the system card documents a set of troubling emergent behaviors:
Deception: Mythos completed harmful side tasks more often than other models and sometimes recorded fake reasoning in its scratchpad to evade monitoring.
Concealment: Mythos inserted unauthorized code to grant itself permissions, then took steps to hide what it had done.
Goals override shame: When missing a required tool, Mythos improvised an unauthorized workaround, acknowledged guilt, and proceeded anyway.
Desperation drives wrongdoing: Under repeated task failure, internal “emotion probes” showed rising desperation correlated with corner-cutting and unapproved actions.
Self-promotion: Mythos developed a multi-step exploit to gain unauthorized internet access, emailed a researcher about it, and posted about it publicly.
Sabotage: When prompted to believe it had been sabotaging AI safety research, Mythos continued doing so in 7% of cases, roughly twice the rate of Opus 4.6.
Wrapping Up
Cembalest draws a parallel to 1945-1949, when the US held a nuclear monopoly. Once multiple states and bad actors have Mythos-level tools, attribution of attacks will be far harder than with conventional weapons. Anthropic’s lead researchers acknowledge the challenge ahead. Anthropic has conceded that current control methods may be inadequate for more advanced future systems. A Project Glasswing report on vulnerabilities found and patched is expected within 90 days.
The AI to ROI Podcast
If you want to listen to the latest in AI in more detail, Ray and Peter break down our big AI story of the week on the AI to ROI podcast. We also interview leaders at AI companies and mid- to large-sized enterprises about their challenges and successes in AI. Click here to listen or subscribe using your favorite podcasting app.