Manus Raises $500M as AI Agent Race Tests Capital and Trust

Manus returns with a record round

Manus, the AI agent developer, has raised $500 million. It is the company's first funding round since Beijing forced Meta to unwind its roughly $2 billion acquisition of the startup 6. The new capital puts Manus back on its feet after a deal that collapsed for political rather than commercial reasons.

The make-up of the investor list says as much as the dollar figure. Every participant is a Chinese institution: Boyu Capital, IDG Capital, Tencent, HSG and ZhenFund 7. No Western fund stepped back in after Meta's exit 7. BigGo Finance describes the round as the largest single raise ever recorded in China's AI application sector 7. That suggests Beijing's intervention redirected money toward agent companies rather than scaring it off.

The same National Development and Reform Commission framework that undid the Meta deal still governs any future exit. Analysts therefore expect the all-domestic cap table to steer Manus toward a Hong Kong listing rather than a US one 7. Manus's story is now as much about the nationality of capital as about its agent technology. Cross-border AI deals involving Chinese startups appear to be closing, and founders may have to plan around that from the first term sheet.

Meanwhile, agents are producing manuscripts at scale

At the other end of the agent landscape, OpenAI is showing what large swarms of AI agents can produce, and how hard that output is to trust.

On October 6, 2026, OpenAI posted 722 mathematics manuscripts to a public GitHub repository under an Apache-2.0 license 12. They were grouped into 372 families of related results 1. All were credited to an internal frontier model that has no name, no model card, no pricing and no API access 4. OpenAI fed the system roughly 4,000 problems and selected the results it judged significant enough to publish 3. The topics span abstract algebra, mathematical physics, quantum algorithms and differential geometry, and some problems are more than a century old 12.

The release builds on OpenAI's September 8 claim that its AI had solved the Navier–Stokes Millennium Prize Problem, with a Lean formal proof 5. According to OpenAI, about 10,000 coordinating agents worked on that problem for 88 hours, using tools including a cached copy of the internet and code execution 11. It arrived in a busy stretch for the company, which also introduced GPT-6.1 Sol and published a system card for GPT-6 Astra. OpenAI says Astra is its first model to reach the Critical cybersecurity tier under its Preparedness Framework 5.

The verification gap

Coverage agrees on the basic facts but diverges on emphasis. Some outlets frame the drop as a transparency problem. Only ten manuscripts include even an abridged summary of the model's reasoning 1. Critics also see a proprietary lab cherry-picking outputs from a model nobody can inspect 3.

Others focus on rigor. Only 162 manuscripts, about 22%, have fully formalized Lean proofs, another 73 are partially formalized, and 487 rest on OpenAI's internal judgment 2. OpenAI's own README concedes that unformalized results "could have issues" 4.

Those concerns quickly proved concrete. A day after release, OpenAI withdrew three manuscripts after a sign error broke one argument and two papers that depended on it, cutting the count to 719 9. A report also found the release departed from guidelines an independent panel of mathematicians had written for AI labs 9.

More pointedly, Alexander Bastounis, Fabian Circelli and Anders Hansen posted an arXiv paper arguing that the Lean code for the Navier–Stokes proof does not match the English argument it was meant to formalize 10. They do not claim the proof is wrong. Their point is that machine checking cannot substitute for peer review 9. NYU mathematician Tristan Buckmaster had questioned the original claim as well 11.

The reaction from mathematicians has been mixed. KAIST professor Baik Hyung-ryeol said several of the problems are of Fields Medal caliber and that verified results would have an enormous impact 8. Twenty-five Fields Medalists signed an open letter, and Terence Tao warned about Goodhart's Law 2. Tao also criticized problems being "solved" by prompters who cannot explain the work 9.

Reading the moment

The two stories describe the same industry from different angles. Manus shows that money for AI agents is abundant but increasingly sorted by geopolitics. OpenAI shows that agent output can now outpace the institutions built to check it.

In my reading, the harder constraint is the second one. Capital will find its routes, whether through Hong Kong or anywhere else. Credibility cannot be raised in a funding round. Until AI-generated claims arrive with formalizations that match their prose and reasoning that outsiders can audit, sheer volume will draw more suspicion than praise.