A Friday with an open-weights model, a warning about numbers that AI makes up and an experiment with 50 agents that turned into a small city hall.

Anyone who works in product knows that the question “which model do we use?” almost never has a single answer.

This week Mistral released the weights of Mistral Large 4, a multimodal model of about 1 trillion parameters, with more than 160 languages and three levels of reasoning. The performance numbers come from the company itself, so it is worth waiting for independent tests. Even so, the whole package stands out.

What interested me was the commercial design. The API costs US$ 0.70 per million input tokens and US$ 2.10 for output, and the weights are available for anyone who wants to run them on their own. The license is a modified MIT: you can use, tune and commercialize the model, but anyone offering it as a service needs a commercial agreement with Mistral.

For those building product, this changes the cost conversation. Instead of choosing between paying per use and being held hostage to pricing, a third option appears: host it, tune it with your own data and control where the information runs. Not every team has the structure for that, and the cost of operating it does not show up in the per-token price table. The guide on AI agent cost management and the cost budget template help compare the options in the same math.

It also helps explain why deciding by ranking alone does not work. Mistral itself cites several different tests, each with its own bias. For a real feature, the test that counts is yours, with your cases. This is part of AI product management and AI agent evaluation.

I like seeing the range of choices grow. More good options mean more negotiating power and more room to design AI products that make the numbers work.

If you are curious and want to see the full announcement, here is the link.

The rest of the radar

  • Artificial Analysis v5 and the benchmark dilemma — the index changed its composition and now uses private tests; choosing a vendor by looking at a single aggregate number is risky. Interconnects.
  • Agent swarms in Minecraft — with 50 agents, leadership emerged only from list order and the invented roles turned into bureaucracy; designing explicit roles and deliverables matters. Y Combinator Labs.
  • Zombie stats — only 40% of the 1,285 statistics cited by AI assistants led to a study that said the same thing; check the source before taking the number into the business case. Trakkr.
  • The economics of intelligence (DeepMind) — AI’s impact on work depends more on agents coordinating well than on each one being more capable, which points to where to invest in product. DeepMind.
  • XREAL Aura with Android XR — Gemini gets to see what the user sees on US$ 1,099 glasses; a new interaction channel, with expensive hardware and a small audience for now. Google.

Next week there is more, with the same idea: separating what changes the product decision from what is only noise.