Good morning, it's Wednesday, September ninth. In today's briefing we have OpenAI's claimed proof of a Millennium Prize math problem sparking a credit fight with a rival mathematician, Meta launching a shopping-and-payments agent called Muse, and two Anthropic researchers resigning over fears the industry is racing too fast toward superintelligence. First up, today in the big model news; OpenAI OpenAI says an unreleased model, more capable than its own GPT-6 Astra, produced a finite-time blow-up proof for three-dimensional Navier-Stokes with an added forcing term, one of the seven Clay Millennium Prize problems. Roughly ten thousand agents worked the problem for about eighty-eight hours, costing millions of dollars, and OpenAI raced to verify it in Lean within days of hearing rumors that two Millennium problems had already fallen elsewhere. NYU mathematician Tristan Buckmaster, who had spent about a year on related work with Anthropic's Levent Alpöge, alleges his private research leaked into OpenAI's effort. OpenAI says no user data was accessed to produce the proof, though it concedes it can't rule out de-identified usage data having improved its models. OpenAI's Sebastien Bubeck called the allegations false and inflammatory, and OpenAI offered Buckmaster co-authorship while excluding Alpöge for working at a rival lab. The proof is self-reported and self-verified, OpenAI is both prover and sole judge, and the Clay Institute has not ruled on whether a version with an added forcing term counts as solving the original problem. The only outside commentary on record, from developer Simon Willison, skips the math and reads the episode as opportunistic credit-taking. Spending millions of dollars and ten thousand agents in under four days on a rumor is now a lab's normal response, turning who publishes first into a resourcing decision. Separately, independent evaluators are finding Astra's headline claims don't hold up as broadly as OpenAI's own benchmarks suggested, winning narrow categories like web development while trailing on general text and reasoning. LMArena's agent leaderboard has Claude Fable 5.1 leading agentic tasks at roughly fourteen and a half percent, ahead of Astra's twelve and a half and Opus 5's eleven percent, and Scale SEAL's HLE has Fable and Gemini 3.1 Pro essentially tied, with Astra absent from the top ranks entirely. That's the fourth independent data point showing Astra's gains concentrated in narrow tasks rather than general capability, and Fable's lead across these boards is the steadier signal for now. Anthropic Infostealer malware has been hijacking Claude sessions through stolen browser cookies, and Anthropic has been force-signing out affected users and issuing refunds. A new report shows the technique escalating: a compromised session key was used to mint unauthorized Claude Code OAuth tokens instead of hijacking the account directly, and one two-hundred-dollar-a-month Claude Max subscriber watched usage climb from forty-five to fifty-five percent of quota during a window with no active tasks running, even with every integration disconnected. Anthropic's dashboard tracks total usage but not itemized usage, so this kind of theft can run for months undetected. Anthropic suspended the account and refunded part of the bill, but didn't send this victim the standard malware-warning email sent to other affected users, suggesting a second, still-unaddressed attack path worth closing before it spreads. Anthropic researcher Jacob Coxon, who had spent three years on pretraining work at both OpenAI and Anthropic, resigned and said labs are racing straight toward self-improving superintelligence and gambling with people's lives, adding that OpenAI staff haven't deeply internalized the stakes while Anthropic understands the risk but proceeds on the belief that no one else will act responsibly. Anthropic's Alignment Science Lead, Evan Hubinger, backed him on the record, saying he personally believes there's more than a ten percent chance AI causes human extinction within a decade, while still calling current models low-risk. Separately, Anthropic technical staffer Mrinank Sharma resigned around the same time, writing in an open letter that the world is in peril and the time has come to move on. Two named, on-the-record departures alongside a sitting alignment lead publicly quantifying extinction odds is a harder data point than the usual anonymous-researcher framing around lab safety culture, and it sits uneasily next to Anthropic's safety-branded narrative heading into an expected IPO. In local model developments, a new open benchmark gives builders a concrete cutoff for how far quantization can be pushed before it breaks reasoning. Testing a twenty-seven-billion-parameter open model across five precision levels found that four-bit quantization performs almost identically to full precision on graduate-level science questions, instruction-following, and agentic coding tasks, while fitting comfortably on a consumer GPU. Two-bit still works for simpler tasks with modest degradation, but one-bit quantization collapses to near-random guessing on hard reasoning, and longer reasoning chains make it worse, as the model burns its token budget without converging. Four-bit is now a concrete ceiling for compressing an open-weight model before it stops being useful. In the harness, tools and orchestration world; Cognition, the company behind Devin and Windsurf, has raised more than two billion dollars at a forty-eight billion dollar valuation, up from twenty-six billion in May and just over ten billion a year earlier. Run-rate revenue has jumped to nine hundred million, an eighty-three percent leap in four months, and the company is training its own model on open-source foundations to cut its dependence on OpenAI and Anthropic and control inference costs. The valuation holds at roughly fifty-three times run-rate revenue, the same multiple as the prior round, and investors point to Cursor's decision to sell to SpaceX for sixty billion rather than raise independently as evidence that coding agents are a multi-winner market, not a single platform that locks in everyone. That confidence sits oddly next to a signal about where enterprise AI actually gets stuck: deployment, not model quality. Google Cloud and Accenture have launched a joint unit training up to a thousand Accenture engineers to embed directly with clients running Gemini Enterprise, the fifth such forward-deployed-engineer program from a major vendor this year after Microsoft, AWS, Anthropic, and TCS. Accenture's own CEO summed up the client complaint driving it: they were promised AI could do the job and it isn't happening yet. Both companies' shares dipped on the announcement, a sign the market reads services-heavy AI rollouts as margin-diluting even when they're necessary. Meta has launched Muse, a proactive personal agent that browses, negotiates, books travel, turns recipe reels into grocery lists, and pays through Link by Stripe, with each user's agent and data running on an isolated dedicated cloud VM. Meta calls it its biggest consumer push since an eighteen-billion-dollar multistate settlement over social media harms, and says none of Muse's data feeds its ad systems, a claim reviewers on Gizmodo and a heavily upvoted Hacker News thread greeted with skepticism given Meta's ad-business history. The isolated-per-user-VM architecture, the same pattern Anthropic and OpenAI use for enterprise agent sandboxing, is becoming the trust baseline for any agent handling payments. The real bet is whether accumulated preference and payment data builds enough lock-in before a free tier that needs to fund itself forces a change to that no-ads pledge. In other news, DeepMind has opened a petabyte-scale, browser-accessible map of predicted molecular effects for every possible single-letter change across the human genome, thirty times the size of its AlphaFold database and free for non-commercial research use. DeepMind keeps commercial licensing, including for its own drug-discovery arm, behind a paid Google Cloud tier. The move removes compute as a barrier to genomic research for academic labs while keeping the commercial upside in-house. That's the briefing. Have a great day, and don't forget to subscribe.