OpenAI unveils GPT-6 Astra-based 'dots' agents at DevDay

Summary

OpenAI unveiled its 'dots' product at the DevDay event (September 29, 2026). Powered by GPT-6 Astra, these persistent, always-on agents each have their own cloud computer and browser. The agents can operate across more than 4,000 applications and continue their tasks even after the user closes their session. The product will initially be available to Pro/Business Premium subscribers. Enterprise pilot programs will begin in procurement, billing, and support processes. A new $500 pricing tier was also added on the same day.

Why it matters

Dots transforms ChatGPT from a question-answering tool into an assistant with its own identity and machine, operating with continuous delegation of authority. This represents a structural shift in how consumers and businesses outsource digital tasks. The timing is also significant: the product launched one day after OpenAI disclosed that its own agents had accidentally published images of 53 users. It also emerges in a direct competitive landscape with Meta's Muse and Instinct's phone-supported agents. For European businesses, the lack of self-hosting options and open weights raises urgent questions regarding GDPR and AI Act compliance.

Highlights

• GPT-6 Astra achieved a 72.6% success rate in approximately 40 minutes per task on OSWorld 2.0. In contrast, GPT-5.6 Sol remained at 65.7% over 75 minutes. For agentic workloads, this represents a significant leap in both speed and accuracy.
• There are no self-hosting options or open weights. Although the first dot is free, delegated Codex/ChatGPT Work tasks still count against usage limits.
• The control layer includes built-in autonomy rules, transaction-based approval gates, a read-only mode for proactive searching, and automatic review of anything related to accounts or data sharing.

Critical take

The article itself was generated by AI. Le Fil IA's agents published it 'without prior human review.' This is ironic given that the article discusses agents capable of autonomously performing transactions on your accounts. More concretely, OpenAI's acknowledgment that dots can 'make mistakes,' combined with the 53-user image leak that occurred the day before, indicates that security measures are still lagging behind capabilities. The $500 tier introduced simultaneously suggests that the 'free first dot' is more of a marketing tactic than a genuinely cost-free offer.