Welcome to Eye on AI. AI reporter Beatrice Nolan here. In today’s issue:
- AI is reshaping who builds, and how fast.
- The U.S. government allows the limited release of Anthropic’s Mythos model.
- AI isn’t killing jobs—yet.
- And California cuts a deal with Anthropic—despite tensions in D.C.
The big news from the weekend is that Anthropic’s Mythos AI model is back—but only for a few.
On Friday, the Trump administration lifted its two-week block on Anthropic’s Mythos 5 model, clearing it for release to more than 100 U.S. institutions, including major companies and government agencies. Fable 5, the public-facing version of the same model family, remains blocked, with talks over its reinstatement ongoing.
Also on Friday, OpenAI got the Anthropic treatment (well, sort of, anyway) over its newest model. The Trump administration asked OpenAI to limit the release of its GPT-5.6 lineup—comprising three models, Sol, Terra, and Luna—to a small group of government-approved partners, citing the models’ advanced cybersecurity capabilities. OpenAI complied while making clear it does not believe “this kind of government access process should become the long-term default.”
We have the full rundown of the news from this week below, but first: AI is fundamentally reshaping who builds billion-dollar companies, and how fast.
AI is accelerating how quickly companies reach unicorn status
The scale of the AI funding boom has been producing numbers that would have seemed fictional five years ago.
Last month, Anthropic, founded in 2021, became the most valuable private company in history after it raised $65 billion at a $965 billion valuation in May, overtaking OpenAI, which had itself closed a $122 billion round at $852 billion just weeks earlier.
New analysis from venture capital firm Accel, produced in partnership with Dealroom and Revelio Labs and shared with Fortune exclusively, shows that Europe is seeing the same unprecedented speed play out at the startup level.
Of the 86 new unicorns minted in Europe and Israel from 2023 onwards, 20% reached a $1 billion valuation (the threshold for unicorn status) within two years of founding, up from just 5% before the generative AI era. Nearly a third got there in three years or less, against 12% previously. The total number of these ultrafast unicorns has also quadrupled since 2023.
Part of what is driving that speed, according to Matt Robinson, a partner at Accel and cofounder of GoCardless, is that AI is a general-purpose technology—one that can be dropped into almost any sector and immediately create value.
Because the value is clearer, sales cycle times collapse, deal sizes grow, and the time between funding rounds shrinks, he said. In AI, the first company to meaningfully address a given market—whether legal, coding, customer support, or dozens of other verticals—tends to lock in that position, becoming the dominant player others then have to compete against. The window to do that is short, and it is open everywhere at once, which is why the pressure on founders to move fast is so intense, he added.
Another reason companies are moving so much faster in the AI age: They are using their own tech. AI—well implemented— can also mean smaller teams and less overhead, allowing companies to scale faster.
Zhenya Loginov, also a partner at Accel, described the ideal founder in the AI age as a “tinkerer”—someone who is constantly testing new tools, aware of everything emerging, and making sure their whole team operates at the same level. Robinson said the best founding teams are also building on top of the tools—running multiple coding agents at once, automating sales outreach and marketing, and compressing internally what used to take months of engineering time into days.
According to Anton Osika, cofounder and CEO of Lovable, which reached $500 million in annual recurring revenue faster than any European tech company before it, said AI has made it possible for anyone to participate in the software economy in a way that was previously closed to them. That step change in who can build, he said, is what is driving the pace and scale of new companies.
A new kind of founder
There’s also been a shift in who is doing the building. The Accel data shows that founders of Europe’s post-2023 unicorns are twice as likely to have come from Big Tech as their predecessors—23% versus 11%—and twice as likely to hold a doctorate, at 18% versus 9%. Academic founders have also doubled, from 12% to 23%. Microsoft and Alphabet have also overtaken BCG and McKinsey as the most common pipeline for founders.
Loginov said that change reflects the maturation of the European ecosystem and the arrival of large tech campuses in London, Paris, and Zurich, giving a generation of engineers experience on the global technology stage before they go off to build.
The rise of PhD founders is also AI-driven—not because AI companies necessarily require a founder with a doctorate, but because AI has accelerated breakthroughs in robotics, cybersecurity, and autonomous software. Founders working on those frontiers tend to come from labs and universities, and the markets they are opening can be very large, the report said.
Osika said he’s noticed that Big Tech talent is also actively moving to Europe to join companies, including his own. “It’s clear that a lot of top talent is moving here,” he told Fortune, adding that the inflow was proof that category-defining companies could be built on this side of the Atlantic—not just in the U.S., where most of them have historically come from.
With that, here’s more AI news.
Beatrice Nolan
beatrice.nolan@fortune.com
@beafreyanolan
FORTUNE ON AI
OpenAI agrees to stagger rollout of its most powerful model to only Trump-approved customers —Eva Roytburg and Beatrice Nolan
Mozilla President: meet the open source ‘rebel alliance’ that could break Big Tech’s grip on AI —Commentary: Mark Surman
AI spending boom accelerates as Big Tech pours trillions into infrastructure —Sheryl Estrada
AI IN THE NEWS
The U.S. allows the limited release of Anthropic's Mythos model. The Trump administration has allowed Anthropic to share its latest model with a restricted group of about 100 trusted partners, following earlier export controls that forced the company to pull the model offline. The broader restrictions still remain, however, and Washington is continuing to vet who can access advanced AI systems. The move offers some relief for Anthropic, which has been trying to convince the Trump administration that its models are safe for general use. Government officials are trying to balance national security concerns presented by new AI models, with pressure from the industry to move faster. However, the ad-hoc approach to vetting new models and the customers that can use them has left critics saying the result is a patchwork system that gives companies little clarity about what comes next. Read more in Reuters.
Meta tightens internal limits on Claude and Codex over distillation fears. Meta is restricting how its applied AI engineering team can use Claude Code and OpenAI’s Codex while building MetaCode, its in-house coding assistant, according to internal guidelines reviewed by The Information. The company is worried that outputs from rival models could inadvertently seep into its own training data, a practice known as distillation that may violate Claude and Codex usage agreements. Engineers can still use the tools for routine work, like organizing code or building test infrastructure, but a human must review all AI output, and external models are barred from generating programming challenges or flagging bugs used to train MetaCode. The restrictions reflect a broader industry tension, with Anthropic recently accusing Alibaba of large-scale distillation and Elon Musk admitting xAI partly distilled OpenAI’s models. Meta is also working to cut its ballooning AI spending by reducing reliance on costly external tools. Read more in The Information.
Google limits Meta’s Gemini access as AI compute shortage intensifies. Google has capped Meta’s access to its Gemini AI models after Meta requested more computing capacity than Google could supply, according to a report in the Financial Times. The limits have reportedly disrupted some of Meta’s internal AI work and pushed the company to tighten controls on AI usage, including more pressure on staff to conserve tokens. It shows how even the biggest tech companies are running into serious compute constraints as demand for advanced AI tools keeps rising. Google itself is racing to secure more capacity, while Meta is actively trying to reduce reliance on outside models by building up its own systems. Read more in the Financial Times.
Chinese AI closes cybersecurity gap with Anthropic. Chinese AI models have rapidly narrowed the cybersecurity gap with top U.S. systems, according to a Wall Street Journal report. Security researchers found that Zhipu AI’s new GLM-5.2 model performs on par with Anthropic’s flagship Mythos model in some bug-finding scenarios, though its not clear if the model can also build working exploits. Cybersecurity firm Semgrep reported that GLM-5.2 outperformed Anthropic’s Claude Opus 4.8 on certain benchmark tests, and with additional prompting, both models can match Mythos in finding software bugs. Separately, Chinese firm 360 Security Technology unveiled a new tool called Tulongfeng that it says performs on par with Mythos. The developments come as the Trump administration has restricted access to advanced U.S. models over national security concerns, prompting critics to argue the policy may be ceding ground to Chinese competitors. Read more in the Wall Street Journal.
EYE ON AI RESEARCH
AI isn't killing jobs—it's boosting headcount. That, at least, is the finding of a new working paper from Ramp Economics Lab, the research arm of corporate spending platform Ramp, which has published a granular dataset on AI's effect on employment. Previous research has largely had to make do with surveys, theoretical "AI exposure scores," and executive commentary. Now Ramp has used its own firm-level spending data, joined with workforce data from Revelio Labs, to study more than 21,000 U.S. companies. The main finding: firms that invest heavily in AI grow their headcount by 10% in the two years following adoption. Entry-level headcount grows even faster, at 12%.
There are, however, significant caveats. The hiring gains are concentrated among what the researchers call "high-intensity" adopter firms—roughly the top third of companies by per-employee AI spend in the first three months, which works out to around $30 per employee per month at the threshold. Low-intensity adopter firms see no statistically significant change. And even the heavy-spending companies don't see results immediately: the gains don't materialize until six to twelve months after adoption, suggesting it takes time for new workflows to spread through an organization before they affect hiring decisions. AI adoption is also unevenly distributed in the first place—VC-backed firms are more likely to adopt, and to adopt intensely, than legacy tech companies, and there are geographic clusters too, with California-based tech firms ahead of comparable New York ones. That means the hiring gains are concentrated at firms that were already heavier AI users to begin with. (It also means that it is unclear to what extent these firms are hiring because AI is enabling their growth or because they are fast-growing startups that would have been hiring any way, in which case their headcount might actually be expanding less than it would have in the past for a similar jump in sales.) In any case, you can read the full paper here.
AI CALENDAR
July 6-11: International Conference on Machine Learning (ICML), Seoul, South Korea.
July 7-10: AI for Good Summit, Geneva, Switzerland.
Aug. 4-6: Ai4 2026, Las Vegas.
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.
BRAIN FOOD
California cuts a deal with Anthropic. Despite drama in D.C. Governor Gavin Newsom and Anthropic have struck an agreement letting California state agencies and local governments access Claude at a 50% discount, paired with free workforce training and technical support from Anthropic's own engineers. Through the deal, Claude will become the first AI productivity tool available to all state agencies through the California Department of Technology's new Statewide Information Technology Shared Services portal. State workers are already using Claude to cut DMV wait times, streamline Medicaid workflows, and automate cyber defense patching. "AI should not replace the human work of government; it should help our workers move faster, solve problems more effectively, and deliver better results for Californians," Newsom said in a statement.
It's a striking contrast to Anthropic's standing in Washington. Earlier this year, the company clashed with the Pentagon over a contract that would have let the Defense Department deploy Claude for any lawful use; Anthropic pushed for carve-outs against mass surveillance and autonomous weapons deployed without human oversight, which angered the administration, in particular, Defense Secretary Pete Hegseth. In response, the Pentagon labeled Anthropic a "supply-chain risk" and blocked it from working with other defense contractors. Anthropic is currently fighting the designation in court.
In contrast to the high drama in DC, California's CIO told Politico the supply-chain risk designation "just didn't come up" during negotiations.
California's deal builds on Newsom's March executive order requiring AI vendors seeking state contracts to demonstrate responsible practices on bias, civil rights, and misuse prevention, and it's the first major commercial agreement to come out of that framework. The governor, a prominent Trump critic, is also rumored to be eyeing a 2028 presidential run, where AI will most likely be a major issue for voters. For Anthropic, it's good optics. Landing its home state as Claude's largest U.S. public-sector deployment is something the company can point to with every other government client weighing whether the company is too much of a liability to touch, even as the federal government has spent months portraying it as exactly that.
