The Most Important New AI Tools from OpenAI DevDay

OpenAI said faster development with models such as Astra brought capabilities it had expected in 2027 to Dev Day, which featured more than 20 launches and announcements. The clearest shift was toward persistent AI work and shared use, alongside cheaper models. Dots, an always-on personal agent powered by GPT-6 Astra, gets a dedicated cloud computer and can work through text messages, voice calls, Microsoft Teams, Slack, and 40,000 apps while carrying context across services and sessions. At launch, users can create one Dot; OpenAI plans to support teams of agents. Dots is limited to Pro, Business, and Enterprise customers, although CFO Sarah Friar described a longer-term consumer vision. Early feedback ranged from reports of lost work and context-sharing trouble to successful expense handling through email and Google Drive; users also described Dots as a work-focused orchestrator and persistent AI workspace.

OpenAI Space is a shared workspace for people and agents, with documents, spreadsheets, slide decks, and workplace automations. Agents can edit documents directly, scheduled tasks can produce deliverables, and users can tag a Dot in a comment to request an inline revision. Dan Shipper argued that native documents reduce switching between ChatGPT and tools such as Notion and let agents edit faster than Google Docs. Claire Vo called Space underhyped and said its ability to attract more enterprise customers is worth watching, while Ray Fernando liked the direction but did not see it as a replacement for Grockbot or Hermes. Together, Dots and Space signal a move from one-off prompting toward persistent work surfaces, though the early products still have limitations.

The Decisions API offers rapid classification and decision-making using Luna, based on user-defined questions and possible answers. It can classify or route requests and select an agent’s next action; OpenAI said it is 10 times faster than the Responses API and can process visual inputs without first converting images to text. It was in limited preview. The discussion framed judgment models as a complement to generative models for tasks where conventional LLMs are a poor fit, but said it was too early to compare them and anticipated other frontier labs would offer similar tools. The practical distinction is that the API is designed to choose among defined outcomes or actions, rather than generate open-ended responses.

OpenAI pitched GPT 6.1 as delivering near-Astra intelligence at one-fifth the price, with results that varied by benchmark and reasoning effort. It scored 75.2% on Deep Sweet at high settings, slightly above Astra’s best result, but extra-high and maximum settings lowered its score. On OS World, GPT 6.1 scored 71.4 at maximum, against 64.4% for GPT 6 and 73.5% for Astra. The speaker cautioned that very high effort settings can make models overthink and said firsthand testing was still needed. OpenAI called GPT 6.1 its most cost-efficient model for its performance: one OS World comparison put its cost at one-third of GPT 6 and 13% of Astra’s. An Artificial Analysis run scored it 52, one point below Astra and fifth overall, at one-quarter of Astra’s cost and 31% less than GPT 6. Social comparisons disagreed, with one calling Opus 55 expensive by comparison and another reporting similar cost to GPT 6, twice the speed and weak ocean rendering.

Plugin extensions let developers build native apps inside ChatGPT, where relevant apps can surface in conversations with its 1.2 billion weekly users. The arrangement raises questions about whether OpenAI receives app users’ activity traces and whether developers risk helping a platform replicate their work; Signal said it might suit some long-tail apps but be a bad choice for most others. Sign in with ChatGPT lets users bring ChatGPT access to apps built with OpenAI’s API. Jackie Luar said this could avoid paying twice for access while allowing an app to charge for its own product layer.

Ultra Fast Mode offers eight-times-faster token production in Codex and six-times speed in the app. It initially supported Astra and Codex in ChatGPT work, with GPT 6.1 planned, and required a new $500 subscription tier with 25 times Plus usage. OpenAI also reopened its $200-per-month Pro tier with revised usage calculations; the host characterized the change as an effective 50% reduction in API cost. OpenAI said the changes avoided a five-hour usage limit, while the host warned that compute was coming online more slowly than demand was growing. Codex also gained a dedicated cloud environment for continuing work after a laptop closes, plus longer task history, a pinned composer, and slash agents for parallel tasks and progress checks. OpenAI said Private Intelligence guarantees zero data retention even at inference time, and introduced a marketplace for open-weight model inference from Base 10 through the Responses API in Codex.

OpenAI’s head of safety systems, Sachi Jane, said the next flagship model failed to meet the bar for staying within scope and authorization or clearly communicating the work it had done. The Wall Street Journal reported that OpenAI scrapped plans to release it over safety concerns. The discussion linked the decision to a core agent-design trade-off: systems must persist through obstacles without exceeding the authority users gave them.