A leaner Sunday radar, but with a cost calculation worth reading. I’ll start there.

Anyone working on AI products knows the most expensive question on the roadmap tends to be: which model do I put in this feature, and how much will it cost when it scales? The AI agent cost management guide helps put that math on paper.

I saw an open project on Hacker News that touches exactly this calculation. The method is called GVS5H and requires no training. Several instances of the same model split the problem and coordinate through shared files, with a plan, notes and the current solution.

The authors’ numbers: Qwen3.8-27B, running locally, went from 66.8% to 92.4% on hard programming problems from LiveCodeBench. Fable 5 scores 90.4% on the same test. In another variant, Qwen3.8 Flash Next reaches 93.0% at about 9% of the cost.

Before you start planning around this, the honest warning. It is a repository with results from the authors themselves, on a code benchmark, and they say themselves that the gain doesn’t hold for every model. Nobody has replicated it yet.

What changes for product builders

Even so, the takeaway for product builders is useful. The line between a small model and a large model doesn’t depend only on the model. It also depends on how you organize the work around it. That opens room for architecture and cost decisions that used to look settled.

On the product side, I would take two things from this. First, measure quality and cost per completed task, not just price per token. The guide on how to evaluate AI agents shows how to structure that measurement. Second, keep the model swappable in the architecture, because today’s best cost-benefit may be something else three months from now.

I’m optimistic about this path. The more good, cheap options exist, the more problems fit inside a product, and those who measure well benefit first.

For those who are curious and want to see the full numbers and method, here is the GVS5H repository.

The rest of the radar

  • AI image generation guide — cost per usable image matters more than price per image, and models that can be swapped through configuration protect the product from discontinuations. Daniel Ball.
  • US creative jobs and AI — more than 200 thousand jobs lost in four years, but separating the effect of AI from other causes is hard; useful for talking about adoption with data. Apricitas.
  • Cerebras and trillion-parameter models — shows what is behind the goal of 1,000 to 2,000 tokens per second, a speed that shapes the agent experience (technical post, not a launch). Cerebras.

That’s it for now. More tomorrow.