OpenAI's Next Model Isn't Out Yet. It Already Solved a $1M Math Problem.

Morning, {{first name | folks}}! OpenAI's next model tackles a $1M math problem, AI runs physics experiments overnight, a tiny startup matches. DeepSeek for $500K, an AI-designed drug, slows aging, and DeepMind maps 9 billion DNA changes. Let's get into it.

OpenAI says an internal AI model, not yet released and already more capable than GPT-6 Astra, produced a full proof resolving the Navier-Stokes Millennium Prize Problem, one of math's seven hardest open questions, unsolved since 1934. Roughly 10,000 AI agents worked the problem for 88 hours straight, sending 2.7 million messages between them, before a formal, machine-checkable proof was verified 17 hours later. OpenAI isn't claiming the prize money.

There's a real race-against-the-clock story behind it too. OpenAI rushed the effort after hearing rumors an Anthropic researcher and an NYU professor had solved something similar, they reached out, and it turned out to be a related but different problem. OpenAI's giving them credit for that one. The claim itself still needs real scrutiny from mathematicians, but if it holds up, this is a genuinely historic moment for what AI can now do.

Beatriz Yankelevich, a graduate researcher at MIT, connected GPT-5.6 Sol to her lab's quantum computing hardware, and it now runs entire calibration sequences on superconducting qubits on its own, choosing measurement parameters, operating the equipment, analyzing results, and deciding what to try next. "I can have agents running measurements for many hours overnight or while I'm working in the cleanroom," she said. "I can check in from my phone, see what they've done, and steer them if something needs fixing."

It's not flawless though, when the signal got weak or noisy, the AI struggled and needed real researcher guidance. Still, this is the kind of work that used to eat days off of a grad student's life, now compressed into unattended overnight runs, freeing her to actually think instead of babysit equipment.

Magic reports that its training approach is more than 10x more compute-efficient than the best open models from DeepSeek, Kimi, and Nemotron. It matched DeepSeek V4 Pro’s base model for roughly $500K, then scaled the approach to a larger model for about $4M that beat every public open model in its tests. On hard math problems, the model reached a 72% pass rate, around the range of Grok 4.6 and ahead of DeepSeek V4 Pro.

Magic describes itself as “likely the smallest team in the world training trillion-parameter models.” The company says it achieved the results without extra training tricks or shortcuts, using roughly 10x less compute than the models it compared against.

Insilico Medicine tested its AI-designed drug rentosertib in a 12-week clinical trial involving 42 people. The drug was originally developed for idiopathic pulmonary fibrosis, but researchers also tracked biological aging during the trial using six different aging clocks. All six showed a reduction in predicted biological age among people taking the drug, while the placebo group showed little change.

The results are still from a small, early-stage study, but the finding gives researchers another reason to investigate the drug beyond its original use. Rentosertib is already being tested in larger trials for pulmonary fibrosis, while Insilico is now looking at whether the same drug could have a measurable effect on aging and other age-related conditions.

AlphaGenome Atlas Can Now Predict the Impact of 9 Billion DNA Changes

Google DeepMind’s AlphaGenome Atlas can now predict what around 9 billion possible single-letter changes in human DNA could do. The system looks at both coding and non-coding regions and predicts how individual mutations could affect gene expression, RNA splicing, protein production, and other molecular processes. It builds on AlphaGenome, which can analyze DNA sequences up to one million letters long.

The Atlas contains roughly 1 petabyte of data, with every possible single-letter substitution across the human genome scored for its potential biological impact. DeepMind is also releasing a Variant Impact Score to help researchers prioritize mutations for further study. The dataset is available for noncommercial research now, with commercial access planned through Google Cloud.