Today, some people were cutting AI costs in half while others were being sued for stealing its secrets. It is almost a perfect summary of the phase we are in: everything is moving fast, including the disputes.

This week, OpenAI launched the GPT-5.6 family, and the detail that caught my attention most was not the model itself. It was the number: partners such as Notion, Figma, and Cursor are already reporting token-cost reductions of 20% to 60%.

That changes a calculation every AI product manager carries in their head.

For a long time, an embedded AI feature meant high production costs. That limited what could go on the roadmap, especially in financial products, where margins are tight and volumes are high.

When token costs drop this sharply, the equation changes. An agent that helps a customer understand a receivables advance, or summarizes a structured credit transaction in plain language, stops being an expensive luxury and becomes viable at scale.

On the credit product side, this is exactly the kind of calculation that determines whether an idea leaves the drawing board. It is not only “can we build this with AI?” It is “can we build this with AI at the volume the business needs, without cost becoming a problem?”

The model also introduced a mode that coordinates multiple agents working in parallel, opening room for more complex automations rather than just one-off answers.

I get excited by this kind of progress because it pushes the boundary of what is economically viable to automate, especially in structured processes such as those that support receivables and credit. And the cheaper it becomes to build well, the more room there is to think about experience rather than cost alone.

For anyone who wants to review the launch details, here is the link: Read more

The rest of the radar

Meta’s Muse Spark 1.1 — a public API for a multimodal agentic model with 1M tokens of context, adding another vendor option when comparing cost and quality. Read more

U.S. government blocks GPT-5.6 release over security — frontier models now undergo government review before general release, creating timeline risk for teams that depend on them in their roadmap. Read more

OpenAI launches full-duplex voice — GPT-Live-1 speaks and listens at the same time, opening a path for voice to become a primary interface in agentic products. Read more

Ask Ad Manager, Google Ad Manager’s agent — a real-world case of a conversational agent embedded in a complex B2B product and a useful benchmark for agentic UX. Read more

Apple sues OpenAI over trade secrets — the dispute exposes the war for AI talent and IP among big tech companies, a risk to consider in cross-company hiring. Read more

An MCP guide for managers — explains the risks of agents using MCP, including prompt injection, overly broad permissions, and lack of auditability, and offers a practical rollout checklist. Read more

Xiaomi cuts inference costs by up to 7x — caching and routing techniques that reduce the cost of long context and longer agent sessions. Read more

Frugon, a free tool to cut LLM costs — runs locally, analyzes call logs, and identifies where a cheaper model can be used without losing quality. Read more

An agent that learns to use any API without documentation — tackles the lack of ready-made MCP integrations in third-party apps, but raises security risks when access is authenticated and outside the product owner’s direct control. Read more

It was a day with a lot of news and little time standing still. See you tomorrow.