Today’s radar brings a new open model, a voice model that learned just by listening, and a reasoning-cost calculation you can bring down. I’ll start with the model.

Every time a new model ships with a finance benchmark in the announcement, I have the same question as anyone working in product: what does this change in deciding which model goes into my feature?

Moonshot launched Kimi K3.1 Pro and a version called Highspeed, which the company says is two to three times faster. It has 1 million tokens of context, and open weights are promised by November 4. The model is also already available through the API, the app and Kimi Code.

What caught my attention was the focus. Beyond code, the company shows results in financial modeling, sourced research and business operations, and says the hallucination rate on a hard test fell from 23.4% to 8.3%. A word of caution: these are internal tests, and the page itself says this is not the everyday rate.

What changes for product builders

For those building products, the useful news is something else. An open, fast model that is competitive on office tasks changes the conversation about cost and vendor dependency. You can think about running it yourself once the weights are out, or use the existence of the option to negotiate price. The AI agent cost management guide helps put that math on paper.

Benchmarks don’t replace your own evaluation

But lab benchmarks don’t replace evaluation with your own data. In any flow that deals with numbers, I first build a set of real cases and compare the models there, before talking about the roadmap. The guide on how to evaluate AI agents shows how to structure this, and the AI governance guide covers the extra care needed when the flow involves money.

I’m optimistic about this movement. More competition among models tends to make what we can deliver cheaper and faster, and those who measure well come out ahead.

If you’re curious and want the full announcement, here is the official page.

The rest of the radar

  • Dessa, from Block — a full-duplex voice model that runs locally points to voice support with lower latency, and Block itself admits the risks are not yet known. Block.
  • ThinkingCap for Gemma 4 — reasoning tokens are cost and latency; cutting 39% on average without losing accuracy changes a feature’s math (the company’s own measurement). BottleCap AI.
  • Island raises US$ 400 million — investors paying for access control and auditing of who acts inside company systems, a requirement that will reach products with agents. Times of Israel.
  • Corgi — a sign that working with several agents in parallel has gained supervision tools, with status, changed files and token consumption. Corgi.
  • Ask HN: why doesn’t OpenAI open 100 companies? — the discussion shows the doubt about where the value lies, in the model or in the application, and weighs on product positioning. Hacker News.

That’s it for now. More tomorrow.