The story that held my attention today was the announcement of a large open-weight model. It comes with four other reads worth your time, from agents doing science to an AI tutor that works but still struggles with adoption.

What we know about Beam

Whenever a new model comes out, the first question people ask me as a PM is usually: which one is best? Over time, I’ve learned that a more useful question is: how much control do I want?

That’s what came to mind when I read about Beam, Reflection’s first open-weight model. It has 501 billion total parameters, with 23 billion active at a time. Reflection says it trained the model on 23.8 trillion tokens and built it for coding, reasoning, and agentic workflows. The company expects to release the weights, technical report, and developer materials later this month under an Apache 2.0 license. Here’s the full announcement.

There is no price or public API yet. That means we can’t conclude what it will cost in production or how its quality compares for a specific use case.

From choosing an API to deciding what to build or buy

What catches my attention is the shift. Every strong open model changes the conversation from “which API should I subscribe to?” to “what makes sense to build, and what makes sense to buy?”

For product teams, this is no longer just a technical choice. It affects cost per use, data privacy, dependence on a single vendor, and how quickly the team can iterate. That analysis belongs in AI product management. When a use case involves autonomy and tools, the AI agents guide helps clarify which pattern fits.

In regulated areas such as finance, running a model in your own environment may unlock a use case that had been on hold. That depends on the controls and requirements of each case; the AI governance guide helps turn risks into decisions. And when the choice affects cost per task, it helps to have a framework like the one in the AI agent cost guide.

I see this shift with a lot of optimism. More options can make more products viable, create room to experiment, and give teams fewer reasons to leave a good idea on the shelf because of cost.

Set your criteria before the release

My take as someone who works in product: start setting your criteria before the model ships. Which task needs the highest quality? Which can use a smaller, less expensive model? Where must the data stay in-house?

Teams with that yardstick can assess each release. Teams without one get stuck following the hype of the week.

The rest of the radar

  • Cloudflare launches Web Search API (beta) for agents — it makes it easier to give agents access to real-time information without separately contracting and integrating each search provider. See the changelog.
  • Opus 5.5 agents identify candidate magnetic semiconductors — a concrete case of agents running a long, reproducible scientific workflow, and a useful reference for autonomous agents in specialized domains. Read the study.
  • Khanmigo after two years: an AI tutor improves math, but adoption is low — rare evidence that the bottleneck for an AI product with demonstrated impact may be engagement, not access or model capability. Read the study.
  • ChatGPT uses real cartoonists’ signatures on fake New Yorker cartoons — a product and reputational risk in generative features: guardrails against false attribution and the use of creators’ identities are missing. Read the report.

That’s it for today. I’ll be back tomorrow.