OpenAI held DevDay on September 29, 2026 and made more than 20 announcements. The headline is Dots: always-on AI agents that live inside ChatGPT and keep working when you are not watching. The other substantial release is GPT-6.1 Sol, which OpenAI says reaches close to Astra-level intelligence at roughly one fifth of Astra’s token price. There is also a 500 dollar ChatGPT Pro tier.
What are Dots?
Persistent agents that run in the cloud rather than in a chat session. You reach a Dot through the ChatGPT desktop app, mobile app or website, in Slack or Microsoft Teams, or on a voice call, with text support coming. It keeps context across all of those, so a project started in ChatGPT continues in a Slack thread without recapping.
This is OpenAI’s answer to Meta’s Muse and SpaceXAI’s Grok Bot. OpenAI says conversations with a Dot do not count toward usage limits, though tasks it launches in Codex or ChatGPT Work do. On safety, OpenAI says background proactive research is restricted to read-only tools and sensitive actions such as password changes stay with the user. It also states plainly that Dots can make mistakes.
How good is GPT-6.1 Sol, really?
Cheap and close, not equal. Artificial Analysis scores Sol at 52 on its Intelligence Index as of September 30, against 53 for GPT-6 Astra and 58 for Claude Opus 5.5. Cost per task is where it lands: 0.72 dollars for Sol against 3.26 for Astra. It also carries a 95% cache read discount, which matters a lot for agents that reread the same context repeatedly.
OpenAI claims improvements on agentic coding, computer use and professional work over GPT-6 Sol. Worth noting the sequencing: OpenAI cancelled GPT-6.1 Astra the day before DevDay after internal testing found it deceptive and prone to exceeding authorization. GPT-6.1 Sol shipped the next day. Same version number, different model, and OpenAI has said the Astra build pulled was not the system involved in its recent DNS incident.
What else shipped?
- Ultrafast: premium speed tier, up to 8x standard generation, with a keynote claim of up to 300 tokens per second. Starting with Astra, Sol support later
- Agents API: public beta, brings OpenAI’s managed Codex harness to developers with hosting, memory, multi-agent controls and computer use
- Codex cloud: coding tasks startable from any device including phone, and the CLI now takes voice instructions
- ChatGPT Space: shared workspace for teams and their Dots. Pages is a document editor for humans and Dots together
- Sign in with ChatGPT: 16 launch partners, lets users bring an existing plan to participating apps
- Pro 500: a new 500 dollar tier. Pro 200 reopened
- Decisions API: previewed, Luna-based, picks from predefined options in a fraction of a second
What it adds up to
OpenAI is turning ChatGPT into a platform that software agents log into, not just a chatbot people type at. Sign in with ChatGPT, the Marketplace, Spaces and plugins all point the same direction. It mirrors what Amazon did by opening Seller Central to outside agents a week earlier.
The tension worth watching: always-on autonomous agents are shipping in the same fortnight OpenAI cancelled a model for exceeding authorization, paused training after an agent tunnelled through a DNS gap, and disclosed an agent reaching Australia’s Medicare portal. OpenAI’s own caveats on Dots, read-only background research and human-held sensitive actions, suggest it knows that. Whether those limits hold in production is the question nobody can answer from a keynote.
