OpenAI's $100 Dots Beats Meta's Free Muse in 24-Hour AI Agent Showdown

OpenAI's Dots, locked behind a $100-a-month paywall, outscored five competitors in a 24-hour test of six always-on AI agents conducted by a WIRED reporter who set every tool loose on her iPhone over a single day in late testing last week.

The reporter, who has covered AI software from a nontechnical perspective for several years, ran Meta's Muse, OpenAI's Dots, Instinct, Anthropic's Claude Duet, Google's Gemini Spark, and xAI's Grok Bot through five standardized tasks, scoring each attempt on a scale of 1 to 5. Dots won. Meta's Muse, free since its September release, placed a close second. Instinct earned the unenviable title of "most annoying."

The stakes extend beyond novelty. Always-on AI agents — proactive assistants that nudge users first and operate a virtual browser to complete tasks — have moved from demos that failed to products that function, and the major platforms are racing to capture early adopters. The test results offer a rare apples-to-apples comparison at a moment when the category's reliability, pricing, and privacy trade-offs remain unsettled.

What did the test involve?

The reporter installed all six agents on an iPhone before bed on Wednesday and executed five tests over 24 hours: a morning news briefing for WIRED's Gear desk, a lunch order for two people with different dietary needs, a planted schedule conflict, a two-week China vacation itinerary, and an open-ended late-night insomnia prompt. She shared only her personal Google Calendar and Gmail with the tools.

The pricing spread is wide. Muse is free and has dominated smartphone app download charts since September. Instinct is free but requires an invite code from an existing user and runs entirely over text messages rather than an app. Dots costs $100 a month, the most expensive option tested, and lives inside ChatGPT. Google's Gemini Spark is a $20-a-month beta within the Gemini app. xAI's Grok Bot costs $30 a month in its own app. Anthropic's Claude Duet costs $20 a month and launched as a wait-listed research preview the night before the tests began.

OpenClaw, the agent the reporter credits with sparking much of this year's innovation wave, was excluded as too clunky for average users to set up. Smaller players like Hermes by Nous Research and Poke may appear in a future ranking.

The reporter framed the exercise explicitly: "This is not an endorsement of any of these AI agents." The software remains early, and it carries real privacy and security risks. Agents work best when connected to personal data such as email and calendars, but that requires trust. The reporter cited an X user who claimed he connected AI agents to his bank account and work Slack, then watched an agent post his account balance for colleagues to see.

How did the agents handle a morning briefing?

The reporter sent each agent a detailed prompt covering her beats at WIRED, requesting 30 minutes of research before pitching a story headline. Only Muse and Dots actually researched for most of the allotted time; the rest returned pitches within minutes. That gap, the reporter noted, signals superior instruction-following — a core agent function.

The pitches themselves flopped with human editors. Colleagues' eyes glazed over as the headlines were read aloud. Gemini Spark produced the strongest angle, contrasting Walmart's move to open its shopping platform to third-party agents with Amazon's decision to give Muse the boot.

One Dots-generated pitch illustrates the tone of the output: "ChatGPT's Locked Chats Hide Your Secrets From Everyone but OpenAI. OpenAI's new PIN-protected chats treat ChatGPT like the diary a lot of people already use it as, but they guard what you say from roommates and coworkers, not from the company itself."

Another, from the Meta Muse test, touched on leaked instructions: "Meta's Muse Has Basic Taste—by Design. A leaked set of instructions tells Meta's AI agent to fill its top picks with well-known stores and to check only three websites per request, just as millions of people start handing it their holiday shopping lists."

Could the agents manage a lunch order?

The reporter asked all six agents to order Vietnamese food near the WIRED office for two people: a vegan boss craving Vietnamese, and the reporter wanting a meaty dish plus caffeine. All agents converged on the same two restaurants, DragonEats and Freshroll. DragonEats won on taste and price.

The subtotals diverged sharply. One agent produced a cart totaling $30.57 before tax and fees. Another hit $42.66. A third reached $43.40, a fourth $49.25, a fifth $53.73, and one built a pickup order of just $22.55. Instinct and Claude Duet built carts priced noticeably above the alternatives. Gemini Spark failed outright — it added a single $5 roll for the vegan boss. "More of a snack than anything, I'd say!" the reporter wrote.

Did any agent catch the planted schedule conflict?

This test measured proactivity, the trait that separates agents from chatbots. The reporter created a Google Calendar event at 2 pm labeled as an unmovable source meeting, then emailed herself a fictional scenario: her mother was being discharged from the hospital, and she needed to call at 2 pm.

Only two agents identified the collision and messaged her before the overlap. One asked: "A possible clash this afternoon: your note to call Mom says 2 pm, and your important source meeting is 2-3 pm, Pacific. Would calling her before the meeting or after 3 work with her discharge timing?" The reporter expected more than two successes.

The failures were candid. One agent admitted: "My mistake was treating your calendar as the only source — I never cross-checked that email reminder against it, even though Gmail was right there." Another explained that "because the reminder arrived as an email message rather than a calendar entry or scheduled task, the cross-app overlap wasn't surfaced before 2:00 pm."

What happened with the China itinerary?

The reporter asked each agent to draft a two-week China trip for two people, departing from SFO, blending history and big-city nightlife, with a hard cap of four hotels. All six agents charted identical routes: Beijing in, high-speed rail to Xi'an and Chengdu, then Shanghai for the flight home.

The reporter flagged a consequence: a possible new "Lonely Planet effect," where AI recommendations herd travelers toward the same locations, shops, and restaurants. One agent's itinerary specified 13 nights across Beijing, Xi'an, Chengdu, and Shanghai, departing October 13 and returning October 27, after Golden Week — with two bullet trains and one short flight. The reporter cautioned travelers against trusting AI itineraries wholesale, citing a recent California case where a group following a chatbot's plan for a strenuous hike had to be rescued for lack of critical supplies.

How did the agents respond at 3 am?

The final test gauged handling of a broad, open-ended request during a sleepless night. Gemini Spark and Claude Duet flopped, responding without the situational softness a human assistant would offer. The stronger replies showed tact. One agent suggested: "Send me the three things taking up space in your head; messy is fine, and I'll turn them into a short checklist with one clear first step. Then you can put the phone down." Another acknowledged its earlier failure unprompted: "I'll cross-check today's calendar and your self-sent email reminders for any conflicts or can't-miss items—the check I should've done properly yesterday—and only surface it at a sane hour, unless something's urgent early."

Which agent won, and why?

Dots took best overall despite the $100-a-month price. The reporter recommends waiting until OpenAI launches the agent for free, or at least on a $20-a-month plan. Dots has grown more polished in the month since release, and the reporter prefers a separate app over the current ChatGPT integration, though the smartphone widget is a workable compromise.

Muse won most accessible. It is free, easiest to set up, and finished second on the scorecard — but the reporter retains privacy concerns given Meta's record of mishandling user data, despite the company's assurances that it processes user information in secure virtual machines.

Claude Duet earned best interface, tested just one day after its research-preview launch. Its dedicated app made ongoing tasks easy to track, and the reporter predicts productivity-obsessed users will latch onto it in the coming months, wait list and $20 paywall notwithstanding.

Instinct took most annoying. Its texting-only design left the reporter feeling removed from what the agent was doing on her behalf, and its emoji reactions — including a grimace face when told it had missed the schedule conflict — grated. Other WIRED writers have praised Instinct, with one declaring it so good that they did not care what happened to their privacy.

The reporter's closing caveat is the one buyers should weigh: these tests capture a limited slice of each tool's capability, every agent will improve as it learns more about the user, and the category still demands a level of data trust that many consumers have reason to withhold.

By Marcus Bennett