Meta Platforms logo (public domain) via Wikimedia Commons
OpenAI's always-on agents earned an 8.8 task average in an independent review, good for third place among five competitors
OpenAI’s new always-on AI agents, called dots, have had their first serious test outside the company. The results were strong, though not quite first place.
An independent hands-on evaluation in October 2026 gave the dots an average score of 8.8 out of 10 across ten common business tasks. That put them third among five competing agents, behind Claude Cowork, which scored 9.3.
What OpenAI actually shipped #
OpenAI introduced the dots at its DevDay event on September 29, 2026. They run on the company’s GPT-6 Astra model.
Think of a dot as a remote assistant with its own desk. Each one gets a dedicated Linux computer and browser in the cloud, so it can work without borrowing yours.
The pitch is persistence. The dots are built to keep running in the background, watch the apps you connect them to, carry out tasks that take several steps, and learn your preferences over time.
Connectivity is broad. Through plugins, the dots can link up with more than 4,000 applications.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Access started right away for ChatGPT Pro subscribers, a plan priced at $100/month, and for Business Premium users in selected markets. Eligible users get their first dot at no extra charge, while versions built for enterprise and specialized work are still in beta.
How the dots performed under pressure #
The independent review focused on the work agents are supposed to take off your plate. Its ten tasks included email management, research, and spreadsheet handling.
The dots averaged 8.8 out of 10 on task performance. They scored higher on the overall experience, earning 9.2.
Claude Cowork’s 9.3 put it at the top of the comparison. The broader field of rivals includes Meta Muse and Gemini Spark, among others.
OpenAI also published its own safety numbers in a system card. According to that internal testing, the dots blocked all 16,600 simulated email attacks, with zero getting through.
The same system card flagged a weaker spot. When the dots handled chains of five tasks, scope violations, meaning cases where the agent wandered beyond what it was asked to do, occurred 8.6% of the time.
Stretch the chain to ten tasks, and that figure rose to 19.7%.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our