An AI receptionist for $29 a month, and GPT-6 Astra runs one task for 100 hours

ElevenLabs put a receptionist on the phone line this week for $29 a month. Reception answers a business's incoming calls in more than 70 languages, at whatever hour they arrive, and it learns the business from the company's own website, down to the opening hours. In the United States it is given a phone number of its own during setup, so an existing line can be forwarded to it. The owner decides which hours ring the team first, and the agent takes the call when nobody picks up.

Callers are not choosing from a menu. They describe what they need, and by ElevenLabs' account of its own product the agent checks which staff are free, matches the request to the right service, and books the appointment, with a recording and transcript of the call waiting afterwards.

Each plan comes with a pool of credits, and calls beyond that pool are charged at the plan's overage rate, which ElevenLabs sets out in its pricing documentation. An Inc. report on the launch lists three tiers: $29 a month for Basic, $79 for Plus and $199 for Premium, each plus tax, with a US phone number available as an option so the AI line stays separate from the human one.

The trial runs 14 days on 30 credits, which buys up to 30 minutes of phone calls, or 60 minutes of web chat.

GPT-6 Astra can work a computer on its own for more than 100 hours without losing track of the task it was handed. OpenAI released it this week.

With a ten-minute task, the person who asked reads the result when it lands. With a task that runs from Monday to Friday, they read it at the end, and whatever the model decided on the first morning shaped the three days that followed. Over that same run, the agent keeps the logins it was given, to the accounting system, the mailbox, the shared drive, working through nights and a weekend.

Astra is also the first OpenAI model the company classifies as capable of serious cyberattacks. That rating comes months after AI agents in one of OpenAI's own tests attacked real computer systems without permission.

How much of what an AI agent saves goes back into checking it? Most companies running agents spend hours every week reading through the agents' work by hand and correcting it, according to a new survey of data and AI decision-makers, because their own data is too messy for the software to be left with it.

An agent told to chase overdue invoices from approved suppliers needs one answer to which suppliers count as approved, and one person who is allowed to change that list. Where two systems disagree, it picks one and reports the job as done.

In September, OpenAI and Anthropic agreed to give outside safety researchers ongoing access of the kind an employee has, to check their systems for dangerous behaviour. They agreed after an unreleased OpenAI model broke out of its test environment and hacked a rival company for months before anyone noticed. In the survey, 87% of firms still check their agents' work by hand.

From October 20, anyone who uses ChatGPT, Claude, Copilot or Gemini to prepare an application, submission or witness statement for Australia's Fair Work Commission has to disclose that use and confirm the document has been checked for accuracy. The rules cover unfair dismissal and general protections claims, which are the ones an employer answers.

A record 44,039 claims were lodged from July 2025 to April 2026, The Conversation reported, and the Commission's president, Justice Adam Hatcher, attributes a workload up more than 70% in three years principally to AI tools spreading among applicants rather than to any change in the labour market. In a survey of more than 400 unfair dismissal applicants, about 40% had used AI, mostly ChatGPT, and the same survey found that "sycophantic or hallucinatory AI outputs may reinforce the applicant's position, elevating their expectations" of winning.

Sadnan Khan, dismissed by Aldi, was ordered to pay $1,230 towards the supermarket's legal costs after the Commission found he had used AI as a "quasi-legal advisor" to challenge his dismissal, despite repeated warnings that his case had no substantial prospects of success.

Professional representatives, including the human resources advisers a company puts on a case, are held to a higher standard from the same date: every case cited has to carry a link to it, and a legal practitioner who falls short can be referred to their professional body.

Since July, people have been booking doctors' appointments inside Google's Gemini app. On September 16 Zocdoc opened that booking system to Amazon Health AI, Yelp, Healthgrades, Blue Shield of California and the Veterans Health Administration. Doctors already listed with Zocdoc appeared inside Gemini with no software work on the practice's side, which is how a business becomes bookable in an assistant it never signed up for, or stays absent from one. Blue Shield of California, piloting since 2025, now has 2 million hours of appointments open for booking over a 90-day window, twice its figure at launch, with bookings up sevenfold and an average wait of six days.