The FTC Probes Who Answers for Agent Actions, ChatGPT Reportedly Moves Beyond the Chat Window, and Voice Agents Get a Per-Minute Price Tag

Three threads this week: regulators asking who is responsible when an agent acts, chat interfaces that stop being only text, and a voice price that makes costs easy to model. Several details come from press reports and aggregators, and we could not open primary pages from our environment, so we flag what is unconfirmed.

1. The FTC Asks Who Answers for an Agent's Actions

Multiple outlets, including ITV and TechRepublic, report that the Federal Trade Commission is investigating OpenAI, Anthropic and other AI developers over possible consumer risks, citing reporting from The New York Times and Reuters. The agency confirmed the probe but has said little about scope. Reports say it will demand documents and seek testimony from executives, and that it may look at whether safety or capability claims were misleading. The coverage describes a run of agent incidents as context. Treat specifics as press-reported, since they rely on unnamed sources.

The question underneath is simple: if an agent is given a task and causes damage, who is responsible? Reporting says the FTC chair has argued that companies cannot avoid responsibility by treating autonomous systems as independent actors. If you ship agents, assume you will be asked to show what yours did and why.

A minimal answer is an append-only record per tool call:

await audit.append({
  at: new Date().toISOString(),
  user: ctx.userId,
  tool: call.name,
  args: redact(call.args),
  approvedBy: ctx.approval?.by ?? null, // who allowed the irreversible step
  result: summarize(result),
});

2. Chat Is Reportedly Growing an Interface

A roundup this week, and a Gizmodo piece titled "OpenAI Wants ChatGPT to Stop Looking Like a Chatbot", describe an "Intelligent UI" for ChatGPT in which the model generates charts, forms and buttons when it judges them a better fit than text. We could not find an OpenAI primary source for the name or the GPT-6 tie-in, so treat it as reported, not confirmed. The pattern itself is established: third-party kits already stream forms, tables and charts from model output, and a form beats five clarifying questions.

For builders the design rule is to let the model choose from components you define rather than emit arbitrary markup. A small schema keeps output validatable and safe to render:

type UiBlock =
  | { type: 'chart'; kind: 'bar' | 'line'; series: { label: string; value: number }[] }
  | { type: 'form'; fields: { name: string; label: string; kind: 'text' | 'choice' }[] }
  | { type: 'button'; label: string; action: 'follow_up' | 'open_link' };

Validate every block, fall back to plain text when validation fails, and keep button actions to an allowlist. A before-and-after screenshot of a text answer versus a rendered chart would show the gain at a glance.

3. Voice Agents Now Have a Price Per Minute

OpenAI's model page lists GPT-Live 1 at $0.05 per minute for the voice layer, billed per second, with backend model and tool usage billed separately at normal rates. Third-party coverage dates API availability to September 10 and reports turn-taking latency of about 0.8 seconds; we could not confirm either from OpenAI's own page, and none of the sources call it "generally available".

Cost modelling is now arithmetic. At that rate 1,000 minutes of conversation costs $50 for the voice layer alone, before backend calls, tools and telephony. Instrument per-call cost from day one, and track minutes per resolved task, because a voice agent that talks longer costs more even when quality is unchanged.

Conclusion

Responsibility, interface and cost all got more concrete this week. Log what your agents do, constrain what they can render, and price voice by the minute before you promise it. Check the vendor's own pages before acting on anything flagged here as reported.

— Maya