AI progress is moving on several fronts at once. OpenAI published what it describes as a solution to the Navier-Stokes Millennium Prize problem, while also shipping a faster image model and showing an agent operating quantum-lab equipment. Google committed €13 billion to infrastructure in Finland. At the same time, US agencies escalated a dispute over model distillation, and Suno launched music models built through label partnerships. The common thread is that capability now depends as much on verification, infrastructure and rights as it does on raw model performance.
OpenAI publishes a claimed Navier-Stokes solution
OpenAI released a proof and Lean formalization that it says resolve the Navier-Stokes existence and smoothness problem by showing a finite-time singularity. The company said an internal model more capable than GPT-6 Astra coordinated roughly 10,000 agents for the successful effort, which used about 130 billion output tokens. GPT-6 Astra then helped formalize and verify the result in Lean.
This is a major claim, not yet a settled mathematical fact. Independent experts still need to examine whether the analytical argument and formalization satisfy the exact prize conditions, and the Clay Mathematics Institute has its own review process. OpenAI said it does not intend to claim the prize. Even with that caution, releasing both a conventional proof and machine-checkable work makes the announcement a significant test of whether AI systems can produce research that survives expert scrutiny.
Google commits €13 billion to Finnish AI infrastructure
Google announced a €13 billion investment in Finland over the next two years, calling it the company's largest single investment in Europe. The program covers digital infrastructure, clean energy and economic partnerships across Hamina, Kajaani, Muhos and Vaala. Google said construction in 2027 and 2028 would support more than 37,000 jobs and contribute €3.6 billion annually to Finnish GDP during that phase.
The energy plan is as important as the server capacity. Google signed a 22-year agreement supporting the life extension of the Loviisa nuclear plant, added onshore wind contracts and backed a 94-megawatt battery system. The package shows how AI infrastructure competition is becoming a long-term energy and regional-development negotiation, not simply a race to order accelerators.
US agencies accuse Chinese AI companies of industrial-scale distillation
A joint NSA, CISA and FBI advisory accused DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI of extracting billions of tokens through millions of requests to US frontier models since at least late 2024. The agencies described the campaigns as malicious knowledge distillation and said operators used cloud services, aggregators and proxy networks called transfer stations to bypass regional restrictions and safeguards.
The advisory recommends stronger anomaly detection, targeted response changes for suspected extraction and cross-company intelligence sharing. Its significance is broader than the named companies. Distillation is a normal research technique, so enforcement will depend on separating legitimate model development from coordinated access that breaches provider rules. That makes API identity, traffic analysis and contractual controls central parts of frontier-model security.
ChatGPT Images 2.5 adds faster generation and directed editing
OpenAI launched ChatGPT Images 2.5 with claimed improvements in detail, reference-image fidelity and multi-turn editing consistency. The company says latency is up to 50 percent lower than Images 2.0. A new Sketch feature turns drawings into image guidance, while templates and image comments give users more structured editing controls.
OpenAI also released GPT-Image-2.5 Flare and Sunburst through its API. The models arrive as OpenAI says users create more than 3 billion images each week across ChatGPT Images and its API. That scale raises the value of repeatable edits and provenance, so the continued use of C2PA metadata and invisible watermarking matters alongside visual quality.
Suno launches label-backed v6 music models
Suno released three v6 music models developed with licensed material from Warner Music Group, BMG and Believe. The standard v6 and experimental v6-wild models are for paid subscribers, while v6-mini is available free. Suno says the models are faster and more expressive, with plain-language track editing and the ability to make songs from image or video prompts.
The commercial model is the bigger shift. Suno will pay participating labels and publishers when the models are used, though the resulting distribution to artists and songwriters is controlled by rightsholders. The launch offers a live test of whether licensing can produce a repeatable framework for generative music while lawsuits and disagreements with other rights owners continue.
GPT-5.6 Sol runs routine quantum experiments
OpenAI described an MIT experiment connecting Codex and GPT-5.6 Sol to quantum-lab software. Graduate researcher Beatriz Yankelevich used the system to choose measurement parameters, operate an uncalibrated six-qubit chip, analyze results and decide what to try next. With clear signals, it completed standard calibration sequences with little intervention.
The agent struggled more when physical signals were weak or noisy and sometimes needed an experienced researcher. That limitation is instructive. Software agents can already extend into laboratories when workflows are instrumented and bounded, but ambiguous physical evidence still requires human judgment. The practical gain is persistent execution of routine measurements, including overnight runs, rather than replacing scientific interpretation.
What matters today
Today's developments show AI moving from demonstrations into institutions that must verify and govern it. OpenAI's mathematics claim will be judged by outside experts, Google's expansion depends on decades-long energy commitments, and the US distillation warning depends on coordinated monitoring across providers. Suno's launch tests a licensing structure, while the quantum experiment tests how agents behave around real hardware. Capability remains the headline, but evidence, infrastructure, access controls and rights will determine which advances become durable.