The dispute between OpenAI and two mathematicians turns on a practical question: when unpublished work is entered into an AI product, can anyone later determine whether it influenced the model? In this case, the answer appears to be no.
Tristan Buckmaster, a mathematics professor at NYU, and Levent Alpöge, a researcher at Anthropic, spent roughly a year on a private collaboration around the Navier–Stokes equations. They used Codex along the way, including for drafts of their work. By late August, they had three Lean-verified proofs of finite-time blowup and believed they had a path toward the larger problem.
On September 3, Buckmaster learned that OpenAI had heard of the work. In meetings three days later, he says OpenAI representatives told him that an internal model had produced a long proof of blowup for a version of Navier–Stokes. OpenAI says its own effort began on September 1 after rumors that Millennium problems had fallen. The parties disagree about the surrounding discussion, including how much human work went into OpenAI’s result.
Buckmaster’s central question was narrower: had the model been trained on, or otherwise had access to, the drafts he and Alpöge had entered into Codex? He writes that he did not receive an answer. OpenAI says its researchers and agents did not see the pair’s work before it was public and that no specific user data was accessed to solve the problem. It also says it cannot rule out that de-identified data derived from product use helped improve its models.
The strategic issue is bigger than one training record. As intelligence becomes cheaper and more available, the value moves downstream into products, discoveries and operating systems. Model companies will follow it. They won’t remain neutral infrastructure providers.
Anthropic is an early example. It launched Claude Science while also announcing a drug-discovery program focused on neglected diseases. That isn’t evidence of customer-data misuse, but it is evidence that a company can sell a platform to drugmakers while becoming a participant in drug discovery and that changes the relationship to pharmaceutical manufacturing companies.
Frontier models may still be the best tool for a proof search, simulation or design loop. But agents use proprietary context and produce valuable work. The decision isn’t only which model gives the best answer. It is where the company’s data, agent workflow and resulting work product should live.
It continues to drive the case for Sovereign AI to ensure you’re keeping your data protected and you remain competitive with model companies who may turn into competitors in your field. Buckmaster’s case is not proof of wrongdoing. It is a warning about where the market is going. As model suppliers continue moving into more industries directly, ownership of your data and your AI infrastructure becomes more important than just choosing your favorite model.
Model Releases
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. Fable is the broadly available version with cyber and biology safeguards; Mythos is restricted to vetted cyberdefenders and life-sciences researchers.
Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, its new agentic and coding workhorse plus a cybersecurity variant.
Alibaba updated Qwen3.8-Max-0902, a one-million-token snapshot post-trained for coding, agent orchestration and visual analysis.
Meta released Muse Spark 1.3 for longer agentic and coding workflows, then launched Muse, a personal agent for coordinating tasks through connected services.
OpenAI released GPT-6 Astra, its most capable broadly deployed model and first to reach its Critical cyber-capability threshold. It also launched GPT-Image-2.5 Flare and Sunburst, new API image models for faster generation and higher-precision editing.
DeepSeek released V4.1 Flash today, the smallest model in its new architecture family and the company’s new default API model. It combines native image understanding with a one-million-token context window. DeepSeek says it outperforms the larger V4 Pro while running faster and at lower cost.
Partnerships & Deals
Compute, Energy & Infrastructure
Nscale signed a compute agreement with Figure carrying an initial $3.5 billion commitment and plans to expand beyond $6 billion. Nscale will also take an equity stake in Figure. The first GPUs are targeted for the second half of 2027 in Barstow, Texas.
Crusoe raised more than $3 billion at a roughly $30 billion valuation in a round co-led by Atreides Management and Valor Equity Partners, Bloomberg reported. The company builds data centers and cloud infrastructure for AI workloads.
Fluidstack raised $1.5 billion at an $18 billion valuation in a round led by Jane Street, according to Crunchbase. It supplies large-scale compute and data center infrastructure for AI.
Gimlet Labs raised a $300 million Series B at a $3 billion valuation led by Andreessen Horowitz. It’s building an inference cloud that distributes workloads across GPUs and specialized accelerators.
Data, Knowledge & Retrieval
AfterQuery reached a reported $3.2 billion valuation, up from $300 million five months earlier, according to Forbes. The company supplies AI training data built around expert reasoning.
Models, Training & Developer Tools
Nvidia confirmed an agreement to acquire Hugging Face for $12.93 billion, following the unconfirmed reports in our last issue. Nvidia says the model and dataset platform will continue to support competing hardware and cloud providers.
HiddenLayer raised a $100 million Series B led by Delta-v Capital to expand security tools for AI models, agents and workflows. Investors included M12, Morgan Stanley and Booz Allen Ventures.
Enterprise Deployment & Distribution
Wonderful raised a $550 million Series C at a $5 billion valuation led by Insight Partners, with Salesforce participating. The funding will support its platform for building and deploying enterprise agents and applications.
MOZN announced partnerships with Redington, sirar by stc, HPE, Microsoft and Edarat to expand enterprise AI in Saudi Arabia and MENA. The agreements cover regional distribution, cybersecurity, private infrastructure and locally hosted deployments.
Industry Applications & Workflows
HiBob raised $166 million with Salesforce providing most of the capital and Farallon Capital participating. Its HR platform brings workforce and skills data together for employee management, hiring and training decisions.
Physical AI & Robotics
HUMAIN and Applied Intuition signed a long-term commercial partnership targeting thousands of autonomous trucks across Saudi logistics routes by 2030. Applied Intuition will provide the self-driving system and vehicle software, with other industrial applications planned later.
Science & Healthcare
Owkin licensed its K Pro AI research platform and multimodal oncology data to Boehringer Ingelheim. It will also generate new immunology data under the agreement, which builds on a 2025 pilot.
Merge Labs signed a multi-year license for Butterfly Network’s ultrasound-on-chip technology to develop brain-computer interfaces. The agreement includes upfront and milestone payments, hardware commitments and royalties on future commercial sales.
DataHow extended its Series A through a bridge round led by Momenta, with Rockwell Automation participating. The funding will expand real-time connectivity and model-based process control in its biopharmaceutical development and manufacturing platform.