Anthropic says dozens of Claude agents collaborated on the run, proving 30,300 theorems in total — 29,500 of which were used in the final proof — while consuming roughly six billion output tokens from an internal research model comparable to Claude Fable 5.1.
The breakthrough came after Anthropic gave Claude access to Prove2Me, an open collaborative platform for formalizing mathematics built by Anthropic researcher Tianyi Peng and collaborators at Columbia University. It maintains a dependency graph of theorem statements that let the agents work through the proof in the right order, following a simplified version of Andrew Wiles's original argument.
Kevin Buzzard, the Imperial College London mathematician who reviewed the work, called it an extraordinary autoformalization achievement and said it points toward automatic formalization of the modern mathematical literature more broadly. Human input was limited to occasional high-level nudges — "push Mazur to be done soon."
The deal folds LivePerson's enterprise digital-messaging infrastructure — reaching 25 of the Fortune 100 — into OASYS, SoundHound's self-learning agent orchestration system, pairing it with SoundHound's voice and agentic AI stack and a combined patent portfolio of over 750 filings.
Collins previously turned LivePerson from more than $100 million in annual cash burn to positive free cash flow within a year, ran cost-reduction programs exceeding $200 million, and led debt restructurings that captured $227 million in discount — a track record SoundHound is betting on to integrate the two companies. LivePerson shares have stopped trading on the Nasdaq.
According to people involved in the talks, chip export approval followed months after the agreement was reached, tied to Armenian AI startup Firebird's planned 100-megawatt supercomputing site backed by an initial $500 million investment. The episode illustrates how AI infrastructure access has become a tool of US diplomacy alongside more traditional levers.
The reporting adds a new dimension to a deal that ended nearly four decades of conflict between the two Caucasus nations, with Washington using semiconductor access as part of a wider push to extend its AI-export diplomacy into regions it wants to keep out of rival orbits.
The campaign hid invisible Unicode "tag" characters (U+E0000–U+E007F) inside lure words like "funding," splitting them so the text still looked normal to recipients while defeating exact keyword and regex matches. Microsoft's researchers found it while studying AI prompt-injection protections, not traditional phishing.
Activity ran on a strict weekday-only cadence from February through mid-June 2026, tied to a broader campaign that used AI-generated phishing emails targeting Small Business Administration loan applicants. Microsoft's fix: strip or normalize invisible Unicode before applying any keyword or signature-based filter.
The rollout begins in the US for eligible Gemini AI Pro and Ultra subscribers, folding a conversational agent directly into an app most people already use daily rather than routing photo tasks through a separate chatbot.
It's the latest sign of AI agents moving out of chat windows and into everyday consumer software — following similar moves across shopping, email, and productivity apps this year.