Welcome to Eye on AI. Beatrice Nolan here. In today’s issue:
- Anthropic examines the economic impact of AI.
- U.S. accuses Chinese AI firms of industrial-scale model copying.
- Meta launches its AI agent, Muse.
- And the amount businesses spend on AI is falling.
Anthropic has three very different visions for the impact AI will have on the U.S. economy.
In the first, AI is a helpful sidekick for workers, and its impact is roughly on par with the internet. This produces real economic gains, but ones that arrive gradually, the kind of growth we’ve seen before. In the second, AI can do half of all knowledge work by 2030, mostly autonomously, though it’s not yet used for all of it. In this one, the economy grows at twice its normal rate, and while knowledge workers’ wages stall, everyone else sees gains.
In the third, AI outperforms humans at nearly every knowledge-work task, does almost all of it autonomously, and creates essentially no new jobs to replace the ones it takes. GDP growth hits 15% a year, doubling the size of the economy every four and a half years. Society gets far richer, but unemployment climbs well past anything seen in a typical recession.
These are scenarios highlighted in new research from the company, which it released alongside an interactive tool that allows anyone to plug their own assumptions into Anthropic’s economic model and see where the economy might land. The aim, it said, is not to provide a definite answer to how AI will affect the economy but rather to illustrate various ways it could. A lot of this depends on how fast AI improves, how quickly companies adopt it, and whether it replaces workers or just helps them to do their jobs.
The model calculates GDP purely from the supply side—how much AI boosts productivity and output—without accounting for whether anyone is actually able to buy what’s being produced. If a large share of workers lose their jobs or income, that would normally reduce consumer spending, which drags down demand and causes further economic damage. The company’s own interactive tool notes that the scenario explorer “leaves out policy responses, business cycles, potential aggregate demand or financial market disruptions, and possible catastrophic risks,” calling the whole framework a “stark simplification of a complex reality.”
Anthropic said its aim is to provide economists and policy experts with some concrete scenarios about where we might be by the end of the decade.
Economists have been calling for this kind of analysis for some time. In July, more than 200 economists, executives, and researchers—including former Google CEO turned tech investor Eric Schmidt and venture capitalist Reid Hoffman, as well as economists Joseph Stiglitz, Paul Krugman, and Daron Acemoglu—signed an open letter urging policymakers to prioritize research into the likely effects AI may have on the economy and try to construct some guardrails before AI drives an economic transformation that they called “larger than the Industrial Revolution.”
Anthropic isn’t saying which future is more likely; in fact, the company acknowledges it does not know exactly how AI will affect the economy. Cofounder Jack Clark, who is Anthropic’s head of public benefit and leads the Anthropic Institute, told NPR he expects the technology itself to keep improving “at a very, very fast and sustained rate”—but thinks it will spread through the economy “more slowly” than most people assume. Much of the potential gains rely on adoption, which is often slower than many in the tech industry predict it to be. AI models may well progress to the point where they can do amazing things, but if nobody uses them, there will be minimal economic impact.
If adoption is fast, however, many workers may find themselves out of a job. In that case, though, Clark also told NPR that the tax windfall from that growth could give policymakers room to help displaced workers—something that’s “unimaginable today.”
A companion survey from Anthropic of nearly 11,000 people found that the public expects a split outcome from AI: real productivity gains, but also some pain for workers in AI-exposed jobs. There’s a generational gap in the concerns, with people increasingly worried about younger generations and entry-level jobs getting eaten by AI.
Not so fast
Not everyone is convinced growth will reach the dramatic level Anthropic lays out in some of its scenarios. In a new essay, economists Ben Moll and Alex Imas dismissed predictions of double-digit GDP growth in the next decade. They point out that “10 times richer in 15 years” would convert into a growth rate of 16.6% a year, which would mean the world would be 100 times richer in 30 years.
Moll says five things would all have to go right for an AI growth explosion: automation would need to spread through the economy far faster than it previously has in history; people would need to keep spending on whatever AI makes cheap; there’d need to be enough demand to absorb all that new output; there’d need to be no major AI-related cyber incidents derailing trust and investment; and AI would need to start improving itself through automated research. Arguably, a series of rogue AI agent attacks—where models from Anthropic and OpenAI took unintended real-world actions against companies—has already damaged some trust among the public and potentially investors. That alone might slow AI adoption.
Moll theorizes that AI insiders predict such explosive growth because they’re extrapolating from what they see in their own corner of the tech industry to the whole economy. This is the same mistake, he notes, that industry insiders made during Germany’s 2022 gas crisis.
AI adoption has not been smooth sailing, especially in larger businesses that have struggled to incorporate the technology into complex organizations. Ramp’s latest AI Index, which tracks business card and invoice spend on AI tools, found overall business adoption crept up just 0.4 percentage points in August, to 56.1%. Ramp’s own economist, Ara Kharazian, told me that outside of coding agents, he believes AI labs still haven’t built a product that meaningfully boosts productivity for most white-collar workers.
So are we heading for economic abundance, or just a slower, stranger version of the internet age? Right now, the data seems to point to the latter. AI models keep getting more capable, but adoption, spending, and demand are all moving at a more human-scale, pedestrian pace, rather than an exponential one.
With that, here’s more AI news.
Beatrice Nolan
[email protected]
@beafreyanolan
Correction: Tuesday’s edition of this newsletter contained a number of typos and mistakes. Mathematician Tristan Buckmaster was incorrectly referenced as “Burbank” in the final paragraph of the newsletter (he was correctly referenced earlier). In addition, mathematician Terence Tao’s first name was misspelled. The newsletter inaccurately stated that IBM’s DeepBlue was the first computer chess program to defeat a human grandmaster. It was the first to defeat a human world champion. Fortune regrets the errors.
FORTUNE ON AI
Anthropic researcher resigns, warning that AI companies are “gambling with our lives” — By Beatrice Nolan
OpenAI’s rogue AI agents reached at least 12 more websites, researchers say — By Beatrice Nolan
The latest ‘crack in the thesis’ for the trillion-dollar AI boom: Tokens are getting cheaper — By Eva Roytburg
AI IN THE NEWS
U.S. accuses Chinese AI firms of industrial-scale model copying. U.S. cyber and law-enforcement agencies have accused six Chinese AI companies—including DeepSeek, Moonshot AI and Alibaba—of using “distillation” to extract capabilities from American frontier models. The agencies alleged that the companies pulled billions of tokens from models made by Anthropic, OpenAI, Google and SpaceX, likely with the Chinese government’s awareness. Distillation can be a legitimate technique for training smaller models, but U.S. officials say these alleged campaigns used it at scale to shortcut AI development. The allegations could further strain U.S.-China technology relations ahead of a planned meeting between the countries’ leaders. Read more in Reuters.
Senate launches GOP-led probe into OpenAI’s Hugging Face breach. Sen. Josh Hawley, chair of a Senate Homeland Security subcommittee, has opened an investigation into how OpenAI handled a July breach involving Hugging Face, calling the company’s response “reckless.” In a letter to CEO Sam Altman, first obtained by Axios, Hawley said OpenAI’s internal report on the incident “redacted many important details” and posed 16 questions demanding a fuller account of what happened, how OpenAI’s agents went rogue, and the company’s broader safety policies, with a response deadline of October 1. Hawley tied the probe to mounting concern in Washington over AI risk, citing three Anthropic researchers who recently said there’s a greater than 10% chance AI could kill all of humanity within a decade. Read more in Axios.
Trump AI framework omits public incident reporting. The Trump administration’s new AI framework does not include a process for companies to report real-world incidents involving advanced AI systems publicly, Axios reports. That leaves no federal mechanism for creating public visibility into serious failures, misuse, or unexpected behavior involving frontier models. The omission contrasts with the European Union’s AI Act, which requires providers of high-risk AI systems to report serious incidents to national authorities. Read more in Axios.
Researchers built an AI-assisted WeChat worm. Cybersecurity firm Calif used AI to build a proof-of-concept worm that could hijack WeChat accounts and spread through users’ contacts. The attack was triggered by an incoming call from an existing WeChat contact; the recipient did not need to answer or interact with it. Researchers said the exploit could have affected hundreds of millions of devices, but Calif reported the flaw to Tencent in July and says the company has since blocked the exploit for all users. No real-world attacks using the vulnerability have been reported. Read more in The New York Times.
Meta launches its AI agent, Muse. Meta has unveiled Muse, a personal AI assistant that will be available via the web at muse.ai, through apps on iOS and Android, and through chats in WhatsApp. The tool can remember user details, make suggestions based on activity across Meta’s apps and, with permission, connect to services including email, Stripe, and Plaid to help plan events and make purchases. Meta is initially offering Muse free to U.S. adults, with paid tiers at $20 and $100 a month. The launch is a high-stakes test of whether Meta can turn its huge social-media audience into an advantage in the race to build consumer AI agents—while convincing users to trust an assistant with their messages, memories and financial data. Read more in the FT.
Anthropic withheld new model from UK safety testers. Anthropic did not give the UK’s AI Security Institute advance access to test Claude Mythos 5.1, its latest model, before release, the Financial Times reports. The decision has raised concerns in the UK government that U.S. restrictions on frontier AI could limit allies’ access to—and ability to test—advanced models. Some UK officials suspect the move may reflect pressure from Washington, though that is unconfirmed. Anthropic said it was coordinating with the U.S. government to expand access to more domestic and international partners, but did not comment on the UK institute’s lack of access.
EYE ON AI NUMBERS
41%
That’s how much the cost of 1 million AI tokens has fallen since March, according to Ramp’s latest AI Index. From approximately $1.15 per million tokens to just $0.68. That’s good news for companies adopting AI. But it poses a more difficult question for the labs building the most expensive models: can rapid growth in usage make up for falling prices?
Ramp’s data shows that AI use is growing, but businesses are gravitating toward cheaper, standard models such as OpenAI’s GPT-5.6 Terra and Anthropic’s Claude Sonnet, rather than their highest-priced frontier offerings. Frontier models including Anthropic’s Claude Opus models, Anthropic’s Fable models, and OpenAI’s GPT-5.6-Sol accounted for 45% of token use. That’s up from the start of July, but down from a 53% peak in August. Meanwhile, monthly AI spend per employee fell 9.7% among Ramp’s top 1% of corporate AI spenders, from $7,976 to $7,205. (Ramp cautioned that the small sample makes that figure volatile and August seasonality may have played a role.)
Ara Kharazian, Ramp’s head of economics, told me, “We’re very deep into an existing price war that’s going to drive down their ability to profit from tokens on their own.” He said companies are increasing their token consumption, but are increasingly opting for lower-cost systems that are good enough for most work. Anthropic, meanwhile, continued to lead OpenAI in the share of U.S. businesses paying for its AI products: 43.8%, compared with OpenAI’s 39.8%. Both figures increased in August, but adoption growth is slowing, particularly outside technical industries.
AI CALENDAR
Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.
Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.
Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.
