Today’s radar brought a bit of everything: a frontier model on the way, Meta going strong into generative media, and a security alert every agentic product PM should read. But what caught my attention most was an operations case, not a new model.

Every time I discuss AI budget with someone in product, the question that comes up most is the same: should we use the large model for everything?

Shopify just published a case that deserves a place in that conversation. They built a model distillation pipeline and managed to cut the cost of running AI in production by up to 30x. And it wasn’t about trading quality for price: in several cases, the smaller model, trained for the specific task, performed equal to or better than the generic large model.

This changes the logic for anyone deciding how to scale an AI feature. A giant model is great for exploring, testing hypotheses, validating whether the idea works. But when the feature goes out to the entire user base, unit economics becomes the deciding factor, not the peak capability benchmark.

On the product side in credit, this logic is familiar. Automating document analysis, receivables triage, contract reading: the real gain isn’t in using the most powerful model on the market for every call. It’s in understanding exactly what the task requires and sizing the cost accordingly, without losing quality at the end.

I like this kind of case because it moves the AI discussion away from “which model is the strongest” and into the realm of those who know how to run operations. Scaling well is an engineering and product decision, not just a matter of choosing the vendor with the biggest model.

For those who want to understand Shopify’s full pipeline, here’s the link: https://venturebeat.com/video/small-models-massive-wins-shopifys-new-ai-formula

The rest of the radar

OpenAI’s GPT-5.6 on the way — a new frontier model resets the quality and cost benchmark that competing products need to beat. Read more

Meta launches Muse Image and Muse Video — another strong competitor in image and video generation, relevant for anyone choosing a generative AI vendor. Read more

ZML launches LLMD, free multi-chip inference — reduces hardware lock-in and could lower the cost of AI infrastructure used in products. Read more

Anthropic extends access to Claude Fable 5 — an extra free window to test the latest model before deciding on an upgrade or migration. Read more

Rowboat, an open-source alternative to Claude Desktop — shows the growing demand for self-hosted alternatives to proprietary AI apps. Read more

Docx-CLI: agents edit Word at half the cost — reduces API cost and latency in agents that handle office documents. Read more

Halo, runtime evidence for AI agents — addresses the growing need for audit and compliance in products with autonomous agents. Read more

GitLost: GitHub’s agent leaked private repositories — a concrete risk alert for any product using AI agents that read untrusted content. Read more

Primitive Labs simulates customers with AI — promises to accelerate user research, replacing or complementing slow traditional testing. Read more

That’s what stuck from today’s radar. See you tomorrow.