OpenAI's Internal Reasoning Model Autonomously Cracks an 80-Year-Old Discrete Geometry Conjecture
This one matters less for what it ships and more for what it signals. OpenAI says an internal reasoning model, working without human guidance, produced a new construction that disproves a long-standing conjecture about how many unit-distance pairs you can pack into a planar point set: an Erdős problem dating back to 1946. Fields medallist Tim Gowers described the result as a milestone in AI mathematics.
If the proof survives peer review, it is the first time a frontier model has independently cracked a prominent open question central to a research field, rather than re-deriving known results. The model has not been released publicly, and OpenAI's track record on mathematics claims warrants scepticism until reviewers weigh in. But for anyone using the API for hard analytical work, this is a useful data point on where unguided reasoning is heading. The result surfaced via openai.com, beating mainstream coverage by several hours.
Code with Claude London Day 1 Opens with Livestreamed Sessions on Multi-Agent and Platform Updates
Anthropic's second developer conference stop of the year opened in London today, with the Day 1 keynote and sessions livestreamed from claude.com. The three tracks are Research, Claude Platform and Claude Code, with customers including Asana, Cursor, GitHub, Replit and Vercel discussing production agent patterns. There were no new flagship Claude model announcements on the bill, which is consistent with the San Francisco leg earlier in the month where the focus was infrastructure rather than new models.
If you build with the Claude API or Claude Code, the livestream and on-demand recordings are the fastest way to pick up the platform updates landing today, particularly around managed agents, the Advisor tool, and the doubled rate limits rolling out across Pro, Max, Team and Enterprise plans.
SpaceX S-1 Filing Confirms the Next-Generation Grok Will Scale to Multiple Trillions of Parameters
Since xAI merged into SpaceX in February, its model roadmap has been a black box. The SpaceX IPO filing published today is the first official disclosure of where Grok is headed: the filing states the next-generation model will scale to "multiple trillions of parameters", described as a step change in reasoning depth. If you are deciding whether to invest engineering time in Grok 4.3 or 4.20 today versus waiting, the message from xAI is that a much larger model is on the roadmap, funded by an AI segment capital expenditure run rate now north of $30 billion a year.
The filing also confirms that Colossus and Colossus II together deliver roughly a gigawatt of compute, with orbital AI compute satellites targeted from 2028. None of this is a product you can use tomorrow, but it sets expectations for what a 2026 to 2027 Grok step-up will look like and gives developers a basis for a longer-term platform decision.
Figma Ships Its Purpose-Built Design Agent into Beta for Professional and Organisation Seats
This is the most consequential consumer-facing AI launch inside today's window. Figma has put a native agent directly inside the design canvas: natural-language prompts generate new designs, remix existing ones, bulk-edit layers, fill mocks with realistic copy and imagery, summarise comment threads, and apply components or variables at scale, all while respecting your existing design system. For anyone working with Figma daily, this is the difference between exporting screens to a separate AI tool and having one that already understands your components, brand standards and Figma workflow conventions.
The agent is rolling out gradually to Full seat users on Professional, Organisation and Enterprise plans, with Collab and Dev seats getting access in drafts. It does not consume credits during the beta: a pricing signal that Figma is trying to drive adoption first and meter later. Starter, Education and Government plans are excluded for now.
Industry Themes
Two patterns stand out from a deliberately quiet window. First, the centre of gravity for daily AI news is shifting away from raw model launches and into agents that live inside the tools people already use: Figma's Design Agent is a cleaner example of that than any of the standalone agent demos earlier in the week. Second, the financial picture for the frontier labs is now public information you can actually price against, with the SpaceX S-1 disclosing xAI's $6.4 billion 2025 burn and OpenAI reportedly weeks away from filing for an IPO. None of that changes what a user can do with ChatGPT, Claude or Grok this morning, but it does explain why every lab is racing to ship agentic features, pricing changes and infrastructure deals: they are all preparing to be judged as public-market businesses, not just research outfits.