OpenAI’s GPT-5.6 Sol proves 50-year-old cycle double cover conjecture

A model proving a fifty-year-old open conjecture in graph theory is the kind of headline that should stop you mid-scroll, and yet by the time you read this it's probably already three positions down someone's feed, wedged between a product launch and a benchmark chart. OpenAI's GPT-5.6 Sol reportedly cracked the cycle double cover conjecture, a problem that has resisted mathematicians since the 1970s. I want to be careful here, because "AI solves famous math problem" has become its own genre of overstatement, and I'd want to see the proof scrutinized by actual graph theorists before treating this as settled. But if it holds up, it's a genuinely different category of achievement than what we usually get from these announcements — this isn't a model summarizing a paper or finding a clever shortcut on a benchmark, it's original mathematical contribution. That distinction matters because it starts to answer the question of whether frontier models can do research, not just describe it.

Meanwhile, the more consequential story for most people building with AI this week is happening at the infrastructure layer, not the frontier-capability layer. The Linux Foundation has stood up three separate governance bodies in six weeks covering AI agent token costs, compliance, and payments. That's a striking pace for standards-setting, and it tells you something: the agentic AI market has moved past the "can agents do useful work" question and into "how do we let millions of agents transact with each other safely and cheaply" territory. When Cohere's VP talks at VB Transform about enterprise sovereignty requiring control of the full agent stack — not just the model, but the data, orchestration, and payment rails around it — that's the same underlying anxiety. Enterprises don't want to hand a cloud provider their entire operational nervous system, and they're right to worry about that.

Atlassian's Jira updates fit the same pattern from a different angle. Turning Jira into an "orchestration hub" for AI agents rather than just a ticketing system for humans is a sensible bet, because the actual bottleneck in agentic engineering right now isn't model capability, it's coordination — who assigns work to an agent, who reviews it, who's accountable when it breaks something in production. By the way, this is also why the Federal AI AGENT Act is worth watching even though it reads like standard consumer-protection boilerplate on the surface. Once agents start making purchases and signing up for services on a person's behalf, questions about consent, liability, and platform interoperability stop being theoretical and start being FTC enforcement priorities.

There's also something worth noting in Soofi S, the 30B open model out of a German consortium trained entirely on Deutsche Telekom's own infrastructure in Munich. It's a small story next to a math proof, but it's arguably the more strategically important one for Europe — sovereignty isn't just a talking point at conferences, it's an actual model, running on actual European cloud, that top's benchmarks without depending on a US hyperscaler. Whether that model finds real adoption outside Germany is the open question I'd be watching next.