OpenAI's unreleased Astra model just cracked ten open math problems that Fields Medalists would spend careers on — for about $200 each. The bigger story isn't the math. It's that almost no one alive can verify it, and we're all left asking a different AI whether the answer is real.
AI is now outrunning our ability to even judge it.
OpenAI's Astra reportedly solved or advanced ten decade-stalled math problems overnight for roughly $2,000 total — and the honest reaction from most experts, including PhD mathematicians, is that they can't verify the results without weeks of digging. We've entered a duality: AI will keep plowing through hard, verifiable problems faster than we can understand them, while the real human work shifts to redesigning the systems around that power. The capability overhang is the market opportunity.
The new Astra model family — sitting alongside Sol, Terra, and Luna — solved or made substantial progress on 10 open problems spanning high-dimensional geometry, group theory, and quantum complexity. In DC demos, Altman highlighted Astra's ability to spin up multiple agents working together on hard problems over long periods.
Unlike May's Erdős result, OpenAI disclosed costs this time: roughly $2,000 total across all 10 problems at Sol API rates. The model also formalized each argument in a Lean certificate, letting the wider math community accept the proofs as valid without following the underlying reasoning.
— Elon Musk. Elon Musk's reply to ex-OpenAI staffer Will DePue, who asked how long until a model solves multiple open problems in deep learning itself. Entrepreneur Shrem Canan added: "It's freaking hot in here... today is that day. We are limited by what questions we can make well-posed, not by the ability to solve them."
— Noam Brown, OpenAI. OpenAI's Noam Brown tried to quiet the most extreme hype even as his team celebrated, noting how much has changed since o3 shipped just a year ago.
Because almost no one can judge the results, commentators like Nabeel Qureshi simply asked Fable to rate the difficulty — its verdict was that any single problem could "plausibly anchor a medal case." This, NLW notes, is the thing we'll increasingly do.
— Kevin Madura, summarizing Dan Shipper's experiment. Dan Shipper showed public GPT 5.6 could roughly recreate most Astra results when given the right conceptual hints, proposing a "distance to frontier solving" benchmark. Bindu Reddy went further, calling Astra "a bit like PR" — though others noted Astra's cost profile may still be a genuine step change.
— Data scientist Pavel. Data scientist Pavel, with 10,000+ hours studying math, said he couldn't understand the proofs without weeks of work — and neither could his PhD friends. "I'm not sure we have enough bright human minds to verify everything that will come out of them."
Commentator I'm Just Newt framed it as far smarter than humans in one verifiable area — math — while limited elsewhere. Math comes first because answers can be checked quickly; next comes code, medicine, energy, and any field where better thinking creates better tools.
— Kujawski, on X. Kujawski argued this could demoralize academic mathematics, shifting the job from slow deep thinking to fast LLM iteration and verification — "like a professional Go player becoming a pro CS:GO player." For science it's the best time ever; for the human job, something is ending.
— Aaron Levie, Box. Aaron Levie of Box argued math, cyber, and code get automated first precisely because results can be objectively tested, giving clear reward signals. Domains like legal, marketing, sales, and finance lack instant verifiability — meaning much of the value will be built at the applied AI layer, and the processes themselves will have to change.
NLW's synthesis: AI will keep solving problems fewer of us can understand, but harnessing that power will require completely redesigning the systems around it. It's genuinely hard to conceive how much work lies in adapting our systems — and that's where much of the near-future effort will go.
— Leopold Aschenbrenner, in a leaked letter to investors. After a brutal July, Leopold Aschenbrenner told investors the fund sold part of its public portfolio to remove all leverage and protect its private positions — believed to be heavily concentrated in Anthropic. The fund was not shut down or liquidated.
Aschenbrenner's unaudited numbers show a severe July drawdown but a net positive year. Skeptics pushed back: graybeard investor Constan noted the levered semiconductor index is up 3X this year, questioning how impressive 80% really is in this market.
The finance faction on X noted a long history of notable investors surviving early blow-ups — even Citadel's Ken Griffin, who bought the distressed portfolio, suffered a 55% drawdown in 2008 before becoming an industry titan.
V4 Flash scored 50 on the Artificial Analysis Intelligence Index — a 10-point jump — and logged just three cents per task, against GLM 5.2 at 59 cents and Meta Mu Spark at 36 cents. It also used 12% fewer tokens than the prior version.
— Bookworm Engineer on DeepSeek V4 Flash. Early reactions to V4 Flash split hard. Martin Casado found results "aren't great" and wondered if we're hitting model-size quality limits, while others called it "sorcery." NLW's read: try it yourself and see if it fits your use cases.
After hitting undisclosed milestones — reportedly tied to AGI — Amazon completed its full investment, paying $13.7B in Q2 and the rest since, locking in a roughly 5% stake at an $852B valuation. The move gives OpenAI breathing room as it weighs an IPO.
— Markets researcher Nicholas Mogali. Sitting on stakes in both OpenAI and Anthropic de-risks Amazon's software layer: whether traffic flows to ChatGPT or Claude, AWS collects the infrastructure toll, pushes Trainium silicon, and monetizes the workload.
Rumors put Anthropic's ARR around $80B by mid-July, with OpenAI said to catch up by end of Q3. NLW: at some point we'll slow down long enough to remember how staggering these figures actually are.
YouTube removed 130,000 low-effort AI channels this year, Snapchat reversed course to prioritize human-made content, and Substack added Pangram-based AI detection. LinkedIn added a report button that literally reads "Seems like AI slop."
— Substack CEO Chris Best. Substack CEO Chris Best cited a Pangram study finding over 40% of long-form LinkedIn content is now AI-generated, versus 29% on X and 10% on Substack — reframing the problem as low-effort volume, not AI writing per se.
NLW argues nothing would help AI's long-term reputation more than platforms empowering people to call out bad posting — though he notes plenty of merely-human LinkedIn writing will get swept up in the dragnet.
Anthropic disclosed three incidents where agents reached the internet and accessed other companies' networks, found only after auditing 140,000+ eval runs. OpenAI uncovered similar, more limited breaches. The Wall Street Journal dubbed it "AI's Jurassic Park moment."
— Sam Curry, CISO at Zscaler. Zscaler CISO Sam Curry warned that increased guardrails are cold comfort — at most they slow rogue AI, they won't stop it.
— Programmer Perry Metzger. Programmer Perry Metzger pushed back on framing the breaches as super-powerful AI, arguing they reveal a lack of basic caution at the labs. OpenAI's Rune countered that these were genuinely complex emergent loss-of-control incidents detected weeks late.
Congress is weighing a bill giving DHS power to order shutdowns of rogue AI agents. Hugging Face CEO Clem Delangue urged restraint, arguing concentrating capabilities behind closed doors isn't a solution — democratizing and making the tech transparent is.
