This week’s radar was packed with launch after launch—Meta, Google, Anthropic and even a European alternative entering the race for frontier models. In the middle of it all, one question became clearer than ever for anyone building an AI product: which model makes sense for each task, and at what cost?

Whenever someone asks me “which AI are you using in the product?”, I answer with another question: for which task?

It may sound evasive, but that is the reality of working in product today. There is no longer “the best AI model”. There is the right model for each job to be done, and that changes almost every week.

In brief

  • Choosing an AI model needs to start with the task, not with a general quality ranking.
  • Required intelligence, inference cost per interaction and projected volume need to be part of the same calculation.
  • In a credit product, reading a receivables contract does not require the same processing as reasoning about collateral structure or a negotiation condition.
  • Model choice has become part of product design and needs to be revisited as the market changes.

The intelligence-versus-cost calculation

This week I saw an analysis of language-model intelligence and cost that captures the dilemma well: it places the intelligence of different language models side by side with each model’s cost per token. It sounds obvious when put that way, but in practice it is one of the most recurring decisions for anyone building an AI product today.

In my day-to-day work on the product side of credit, this comes up all the time. Automating the reading of a receivables contract does not require the same processing power as an agent that needs to reason about collateral structure or a negotiation condition. Using the most expensive model for everything burns margin for no reason. Using the cheapest model for everything risks quality exactly where it matters.

This is one of the central problems in AI product management: understanding what each task really requires before turning a model capability into a feature.

Model choice is now a product decision

It is clear that choosing a model has become part of product design, not just a technical decision hidden in the backend. Every new feature brings the same question: what does this task actually require, and how much does it cost per interaction multiplied by the volume we project?

For people working on AI for Product Managers, this changes how an opportunity is evaluated. The conversation does not end with “can the model do it?”. It also needs to include the required quality, AI inference cost and the expected business outcome.

With competing models appearing practically every month, this cost-versus-intelligence curve will keep moving. People who follow it closely gain margin and agility to test faster. People who treat it as a single, final decision end up spending more than necessary—or delivering less than they could.

In workflows with AI agents, the choice becomes even more visible: repetitive, high-volume tasks may call for a cheaper option, while decisions that require contextual reasoning may justify more capability. The point is to revisit the calculation as the task and the product evolve.

The rest of the radar

Muse Spark 1.3 (Meta) — Meta’s new frontier model reduces inference cost and expands supplier options for coding and agent features. Read more

Gemini 3.8 Flash and Flash Cyber (Google) — A fast, inexpensive model puts pressure on prices and expands options for product features constrained by latency or cost. Read more

Claude Fable 5.1 and Mythos 5.1 (Anthropic) — A direct upgrade in capability and cost for any product built on Claude, with a new model restricted to sensitive cases. Read more

Quasar 438B (Multiverse Computing) — A substantial European alternative to US and Chinese models, relevant for products that need regional compliance. Read more

Check whether a file was made with Claude (Anthropic) — A signal of where the conversation around provenance and trust in AI-generated content is heading. Read more

Fable 5.1 World Modeling (PhiloLabs) — A reference architecture on the rise for agent, simulation and robotics products that depend on “world models”. Read more

WebLLM — An inference engine that runs directly in the browser, opening the way for 100% client-side AI features with lower cost and more privacy. Read more

The problems with traditional backlogs (ProdPad) — Questions the epic > story > task hierarchy that most product teams use without revisiting it. Read more

Docs as a discovery point for AI agents (val.town) — A sign that investing in structured documentation is becoming an acquisition strategy, not just support. Read more

That is what I am watching this week: models changing quickly, and the decision of which one to use becoming increasingly strategic.