This week had more frontier model launches than you can properly keep up with — Grok 4.5, GPT-Live, GPT-5.6. But what kept rattling around in my head today wasn’t any of those. It was a much simpler question about price.

Every week someone asks me how much it costs to build an AI product. And I increasingly realize that’s the wrong question.

I read an article from Mercury this week that sums up this point well. The core idea: if your AI product became cheap to build, that doesn’t mean it should be cheap to sell. Price doesn’t come from the cost of running the model. It comes from the value the solution delivers to whoever uses it.

It makes perfect sense. An AI feature that saves ten hours of manual work per week is worth far more than the few cents spent on tokens to deliver it. Charging based on infrastructure cost is leaving money on the table.

On the product side in credit, this logic isn’t new — it just gets an extra chapter with AI. Credit pricing has never been about the cost of processing an operation. It’s always been about the value of unlocking working capital, advancing a receivable, giving cash flow predictability to those who need it. AI changes the cost of delivery; it doesn’t change this yardstick.

What I like about the article’s framework is that it’s highly practical: map the real value generated, look at what the market already pays for similar outcomes, choose between charging by usage, by seat, or by result, test willingness to pay, and adjust. It works just as well for someone launching an AI feature as for anyone designing a financial product.

A reflection for those running AI cheaply inside their product today: cheap to build shouldn’t automatically become a discount on price.

For those who want to dig into the full reasoning, here’s the link to the article: https://www.aol.com/articles/price-ai-product-cost-almost-133004000.html

The rest of the radar

Grok 4.5 — xAI’s new frontier model targets Opus 4.8 and GPT-5.5 with lower prices, pressuring pricing across the entire LLM market. Read more

GPT-Live — new full-duplex voice architecture resets the UX bar for conversational products and paves the way for longer, more agentic voice agents. Read more

GPT-5.6 (Sol, Terra, and Luna) — OpenAI’s public launch arrives in the same week as releases from xAI and Meta, intensifying the race for enterprise adoption. Read more

SWE-1.7 — Cognition’s coding agent with performance close to GPT-5.5 and Opus 4.8 at a much lower cost, relevant for engineering productivity stack decisions. Read more

Signal vs. noise in coding evals — OpenAI’s audit found about 30% of SWE-Bench Pro tasks broken; you can’t blindly trust “capability” rankings when choosing a model. Read more

Flint — Microsoft Research’s visualization language simplifies reliable chart generation by AI agents, reducing dev effort in analytics features. Read more

Frugon — open-source, local tool that identifies which LLM calls can migrate to cheaper models, straight to the point of AI FinOps. Read more

Agent washing — term coined by Gartner warns of products sold as “AI agents” that are in practice just if/else logic with an LLM layer on top. Read more

Agentic Ziggy (Akeneo) — AI agent for product data teams reduces manual catalog work while keeping human supervision over approvals. Read more

That’s what I picked out today. More tomorrow.