OpenAI Takes Aim at Grok and JEV

The story

OpenAI’s latest DevDay was not just another model showcase. It was a platform move. The company introduced Dots, a class of “always-on” agents inside ChatGPT that can take ongoing responsibility for work, maintain context over time, use their own cloud computer, and connect through OpenAI’s plugin ecosystem to more than 4,000 apps [1]. In the same event recap, OpenAI also introduced the Decisions API, a limited-preview interface that uses Luna intelligence to answer user-defined questions with finite, predefined answers, giving developers a faster way to classify, route, or choose an agent’s next action [2].

That combination is why this looks like a two-front attack: Dots put OpenAI in the race with Grok-style always-on assistants, while Decisions API pushes into the JEV category of small, fast, structured decisions. OpenAI is not merely adding features to ChatGPT; it is trying to make ChatGPT the place where persistent agents live, remember, coordinate, and decide [1].

Dots: ChatGPT grows a background process

The key shift is persistence. A normal ChatGPT conversation waits for a user prompt. A Dot is designed to keep moving between conversations: it can monitor a project, prepare work, ask for approval when needed, and bring results back when the user returns [1]. OpenAI describes Dots as agents that learn from feedback, use their own cloud computer and browser, and work toward user goals around the clock [1].

The initial use cases sound deliberately mundane: calendar checks, research, notes, invoices, app testing, launch materials, scientific analysis, sales proposals, content production, and software fixes [1]. That matters. The market for agents will not be won by cinematic demos alone; it will be won by systems that reliably handle the boring connective tissue of digital life.

OpenAI’s pitch is that a Dot can sit across ChatGPT, Slack, and Teams, carrying context from one channel to another instead of forcing users to restate the whole project each time [1]. SiliconANGLE reported that users can message or call their agent in ChatGPT on desktop, web, or mobile, and also reach it in Slack and Teams, with context following the agent between channels [5]. That is the real product surface: not a smarter answer box, but an assistant with continuity.

Why this is aimed at Grok

The Grok angle is about the always-on assistant category. Ahead of the keynote, 9to5Mac framed the expected launch as ChatGPT’s response to SpaceXAI’s Grok Bot and Meta’s Muse, built on existing agent tools but with broader consumer appeal [6]. Axios similarly described Dots as OpenAI’s take on the AI assistant and as part of a wider contest over agents that can book travel, cancel subscriptions, and perform practical tasks [4].

OpenAI’s first distribution choice is telling. Axios reported that Dots are available to Pro, Business, and Enterprise users, with one Dot per user to start, and that OpenAI eventually imagines teams of Dots working together [4]. In other words, OpenAI is not starting with a free mass-market social assistant. It is starting where users already pay for productivity, coding, and business workflows.

That gives OpenAI a different wedge than a consumer chatbot built for personality or viral use. Dots are positioned as work infrastructure: each one can use connected apps, Codex, and ChatGPT Work to research, analyze data, prepare documents, and build software [4]. If Grok-style assistants compete on immediacy and personality, OpenAI is trying to compete on integration, memory, and permissioned action.

The JEV front: decisions, not prose

The second front is quieter but just as important. The Decisions API is aimed at cases where software does not need a paragraph; it needs a choice. OpenAI says the API focuses Luna’s intelligence on a predefined set of possible answers, using text or images as context, and returns answers for classification, routing, or selecting an agent’s next action [2]. It is in limited preview, with a broader release planned in the coming days [2].

That is exactly the area that JEV-style systems made newly interesting: fast, constrained, repeatable judgment. Instead of asking a general model to explain itself at length, the application asks: approve or reject, route to sales or support, escalate or close, use tool A or tool B. The output becomes part of software logic rather than a chat transcript.

Axios noted that the Decisions API has drawn comparisons to TypeFace AI’s Jev and is meant for companies that want AI judgment without giving systems unlimited freedom [4]. That framing is crucial. The next generation of agents will need millions of small decisions, but not every decision should be a full reasoning trace from a frontier model. Some should be cheap, bounded, auditable calls.

Control is the product

Persistent agents create a trust problem. If a Dot can work while a user is away, connect to apps, read background information, and prepare actions, then permissions and safety controls become core features, not compliance footnotes.

OpenAI says each Dot runs in a protected cloud workspace, can use connected accounts chosen by the user, and can optionally be granted access to a personal computer [3]. The company also says proactive research uses read-only tools, so background tasks cannot directly send messages, change connected app content, or control a browser or desktop [3]. For consequential actions, OpenAI describes an Auto-review system that checks planned steps against instructions, custom rules, and safety requirements before a Dot sends emails, changes files, or performs similar actions [3].

The design still requires caution. OpenAI explicitly says Dots can make mistakes, and its safety blog frames the system as a set of layered protections rather than a guarantee [3]. The commercial challenge is obvious: users want agents that do real work, but the more real the work becomes, the more damaging a wrong click, wrong recipient, or misunderstood instruction can be.

Why the 4,000-app number matters

The claim that Dots can connect to more than 4,000 apps through OpenAI’s plugin ecosystem is not just a feature count [1]. It signals that OpenAI wants ChatGPT to become a control plane for other software. Calendars, email, documents, task boards, code repositories, spreadsheets, CRMs, and chat tools are where work already lives. A persistent assistant becomes useful only if it can operate near that existing context.

This is also where commercial viability starts to look realistic. A single agent that drafts generic text is easy to copy. An agent that has persistent context, app permissions, approval rules, workspace controls, Slack or Teams presence, and links into Codex becomes harder to displace. OpenAI is trying to make the assistant less like an app and more like an operating layer.

What to watch next

The immediate questions are practical. How fast will Dots roll out beyond eligible Pro, Business Premium, and Enterprise users? How reliably will they handle messy, multi-day tasks? How many actions will require approval in real use? And how quickly will OpenAI expose pricing and technical details for the Decisions API beyond the limited preview [2]?

The deeper question is whether users want a single persistent AI identity working across their digital lives. If they do, OpenAI has just made its clearest move yet toward that future. Dots are the background workers; Decisions API is the low-latency judgment layer. Together, they show OpenAI aiming not only at Grok-like assistants, but also at the JEV-shaped infrastructure underneath the agent economy.

Sources from the last 72 hours

1. [1]Introducing dotsSep 29, 2026, 2:00 AM
2. [2]DevDay 2026 RecapSep 29, 2026, 2:00 AM
3. [3]OpenAI debuts dots, its assistant to take on MuseSep 29, 2026, 7:15 PM
4. [4]OpenAI launches Dots, always-on AI agents in ChatGPT with their own cloud computersSep 30, 2026, 12:31 AM
5. [5]OpenAI makes 20+ announcements at DevDay event including always-on agents called dots, watch liveSep 29, 2026, 6:03 PM
6. [6]How we build safety, security, and privacy into dotsSep 29, 2026, 2:00 AM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.