Sunday, ten items on the radar, and one of them stayed on my mind all day. It was not a model launch, but a small change on a settings page. I will start there.

In brief

Cursor removed cost information from its usage page and CSV export. The reaction from paying customers was immediate, with hundreds of comments on Hacker News; the issue was not the subscription price, but the loss of visibility. In a consumption-priced AI product, cost transparency is not just reporting: it is a roadmap feature that lets users predict spending before the bill arrives.

There is something I learned while working with credit products: customers can tolerate a high price, but they cannot tolerate not knowing how much they will pay.

I remembered that this week while following the Cursor case. The company removed cost information from its usage page and CSV export. The reaction from paying customers was immediate, with hundreds of comments on Hacker News, and the issue was not the subscription price. It was the loss of visibility.

It makes perfect sense. In a consumption-priced product, users do not buy only the tool; they buy predictability. When you remove the cost dashboard, people are not just annoyed, they become uncertain. And uncertainty is what holds back renewals, usage expansion and internal budget approval.

In finance, this is almost obvious because we have learned it the hard way. There is total effective cost, there are statements, and there are transparency rules precisely because trust in any credit relationship depends on clarity about what is being charged. Anyone who works with receivables and advances knows that the hard conversation is never about the rate itself; it is about the rate the customer did not understand.

Cost transparency is a product feature

The most interesting point is that AI tools are now discovering this in practice. Token-based billing has brought back a model similar to energy consumption, along with the need to give users a trustworthy meter.

For anyone building an AI product, there is a very practical question: can users predict their monthly spend before the bill arrives? If the answer is no, cost transparency has stopped being a report and has become a roadmap feature. It is an AI product management decision: cost, expected outcome and user experience need to be part of the same conversation.

In workflows that use AI agents, that predictability also depends on making clear which steps consume more and which actions can increase usage. In sensitive contexts, AI governance helps define responsibilities and controls; the AI risk matrix helps calibrate limits, human confirmation and audit evidence.

I remain optimistic about the speed of this industry. I just think maturity will come less from the models and more from these unglamorous choices about how to charge, how to explain and how to keep customers in control.

For anyone who wants to see the original discussion behind this, here is the link: Usage page to token amount — what?

The rest of the radar

Dynamic Workflows in Claude Code — the coding agent moves beyond a single prompt and starts chaining steps and decisions, changing how teams design internal automation. Read more

Step 3.7 Flash — another fast, low-cost model is putting pressure on cost per token and reopening features that were discarded as economically unviable. Read more

qm, a multiplayer harness for agents — shows the next frontier: multiple agents and humans working on the same task, with a queue, visibility and an approval point instead of an isolated chat. Read more

Robinhood lets AI agents trade stocks — the first high-risk case of an agent with real execution permission; the patterns for limits, consent and auditing that emerge from it will become a reference. Read more

Noisy LLM evaluators are still useful — removes the excuse that AI cannot be measured: even an imperfect eval provides enough signal to prioritize improvement. Read more

Study points to dark patterns in AI chatbots — defines the ethical and regulatory line that will be demanded of any conversational product, starting with success metrics. Read more

Software for One — questions the SaaS premise: if generating software becomes cheap, generic products lose value compared with tailored ones. Read more

HubSpot opens Agent Hub and Agent Builder in public beta — a major SaaS company turning agents into a native layer redefines what customers start to expect as standard. Read more

Algolia Agent Studio for shopping assistants — the winning e-commerce pattern is an agent grounded in the verified catalog, not free-form generation. Read more

That is it for today. Have a good week, and good work to everyone opening the roadmap tomorrow.