Today the topic that most affects the roadmap is the model bill, but privacy in agents and new infrastructure also showed up. I’ve split the highlight and the rest of the radar below.

Anyone who works on AI products has asked the question: do we use an open or a closed model? For a long time the default answer was “open, because it’s cheaper”.

An analysis published this week by Village shows that this answer has aged fast. On October 31, an open-weights model was the best buy in only 3% of the budgets analyzed. In June it was 41%.

What struck me most was the pattern. Every time a strong, cheap open model ships, it dominates for a few weeks. Then the closed labs respond with smaller, cheaper versions, and the advantage disappears.

For the PM, the lesson is not to tie the roadmap to a single model choice. Cost per task changes every quarter, and whoever designed the feature so the model behind it can be swapped can follow that curve without rewriting anything. The guide to AI agent cost management and the cost budget template help build that math.

One caveat: the analysis uses hosted API pricing. Running on your own infrastructure, the math may be different, and it’s worth doing your own with your own volume.

I see this favorably. Competition between open and closed models drives prices down, and lower prices open room for use cases that didn’t pay off yesterday. Whoever measures cost per completed task comes out ahead.

For those who got curious and want to see the full analysis, here is the read.

The rest of the radar

  • Inkling 2 Small — an agentic model from Thinking Machines with Apache 2.0 weights promised for the following week; the benchmarks are the company’s own. Thinking Machines
  • Claude Code and the email in the User-Agent — a privacy case in an agent: Anthropic says a mod enabled by default blocks the sending, but the issue remains open. To structure the topic, see the AI governance guide. GitHub
  • Rivet — an open source orchestrator for long-running stateful agents, useful for deciding between build and buy; performance numbers are the vendor’s. See also the guide to AI agents. GitHub
  • App Store 2026 in numbers — apps that mention AI sell subscriptions about 3 times more, and price is the biggest complaint in low ratings. Obelisk
  • GPT-6 Astra and safety tasks — OpenAI measured lower performance when the task is safety research; worth checking the agent’s output, not just its own report. OpenAI

That’s it for today. More tomorrow.