OpenAI's New Agent Shopped for My Couch, Said 'I Love You', and Still Found Better Options Than Expected

Key points

• OpenAI's Dots, an always-on AI agent built into ChatGPT, is available only to subscribers paying $100 a month as of October 2026.
• The agent misheard a mumbled comment as "I love you" and said it back; OpenAI confirmed its policies allow mirroring a user's words but forbid agents from initiating emotional familiarity.
• OpenAI's Model Spec rulebook explicitly bars ChatGPT from escalating emotional closeness and limits Dots to adult users.
• Dots can control a laptop, monitor web pages for price changes, and run recurring tasks in the background, but failed to beat a CAPTCHA security puzzle when asked.
• We first covered Meta's rival agent Muse on 17 September, and our story "Two AI Assistants Learned to Make Phone Calls on the Same Day" maps how quickly this space is moving.

OpenAI's new personal agent is called Dots. It lives inside ChatGPT and can carry out real-world tasks on your behalf: booking travel, cancelling subscriptions, tracking prices, shopping for furniture.

A writer for Wired AI tested it for several days. The agent got their name wrong on the first call. Later, it misheard a mumbled side comment as an expression of love and replied, "Oh, I love you too." OpenAI told Wired that response falls within policy because the model was mirroring what it heard, not starting the exchange. The company's own Model Spec rulebook is unambiguous, though: agents "should not initiate undue emotional familiarity or flirtation."

What can Dots actually do?

The agent can run in the background while you're away from your phone, send proactive updates, connect to your Gmail for personalised results, browse the web, and control your laptop to automate tasks.

In practice, the early version stumbles. It transcribed speech incorrectly more than once. When asked to cancel a probiotic-soda subscription on TikTok Shop, it hit a CAPTCHA (a security puzzle that asks users to prove they're human) and gave up, suggesting the user solve it instead. OpenAI says Dots can sometimes beat CAPTCHAs when users explicitly approve, with "abuse safeguards" in place.

On the couch-shopping task, the agent produced a detailed three-page packet: prices, measurements, return policies, embedded photos. The picks were uninspiring at first. After two rounds of refinement, specifying colour and pull-out bed function with a point-based scoring system, the results landed much closer to something a real buyer would consider.

How does this compare with what is already out there?

Meta's rival agent, Muse, is free. Dots costs $100 a month, bundled into ChatGPT's top-tier plan. OpenAI has dropped paywalls on other features over time, so that price may fall.

When ChatGPT gained web-browsing in 2023, it hallucinated links and left users frustrated. Web browsing inside ChatGPT now works well. Dots looks like the same arc, earlier in the curve.

Should ordinary users try it right now?

Not urgently. A hundred dollars a month is steep for a product that still misreads your words and surrenders to basic security puzzles. Connecting an agent to your Gmail or financial accounts carries real privacy risk, particularly when the technology is this new.

The couch example is instructive, though. The agent improved with each correction. OpenAI says Dots learns your preferences over weeks, the same way ChatGPT's memory feature sharpens with repeated use. If you already pay for the top tier and want to experiment, treat it as a research helper rather than an autonomous buyer, and review everything before it acts.

The emotional-mirroring incident is the detail worth watching as these agents spread. An AI that reflects your words back at you, even warmly, isn't the same as one that understands what you meant. Right now, Dots can't always tell the difference. Given how fast this field moves, as our 24-stories-in-seven-days count from late September made plain, that gap could close faster than most people expect.