Oct 3, 2026 · 10 min read · Trending AI

OpenAI Dots Explained: Always-On Agents, Price & the Open-Source Answer

At DevDay on 29 September 2026, Sam Altman called it "more ambitious than ChatGPT": Dots — always-on agents that keep working after you log off, each with its own cloud computer. Two days later, an open-source clone shipped on GitHub claiming the same capabilities at 1/10th the cost. Here is what Dots actually are, what they cost, how they compare to Meta's Muse — and the self-hosted alternatives, with an honest verdict for Indian businesses.

What Dots actually are (DevDay facts)

Dots are personal agents powered by GPT-6 Astra — OpenAI's best coding model to date. You create one "primary dot", give it a name, and hand it goals. It then works across 4,000+ connected apps with its own browser and computer, messaging you with updates and questions like a text conversation, or over a voice call from ChatGPT web, desktop or mobile. They also live inside Slack and Microsoft Teams (context carries across), with SMS texting announced as coming soon. Teams can share a Dot inside ChatGPT Space, a new collaborative workspace.

• Developer example: a Dot watches customer feedback, scopes fixes, builds + tests updates, and returns finished pull requests with a video showing its changes
• Creator example: a Dot reviews an interview transcript, finds clip moments, writes show notes and social posts
• Internal example: at OpenAI itself, Dots start investigating the moment a bug appears in Slack
• Enterprise: "specialist Dots" with their own identity, credentials and system access — tested in procurement, invoicing, email marketing, support and contracting; Microsoft Agent 365 governance integration in progress

How they work: cloud computer + guardrails

The big architectural idea: your computer and the Dot's computer are separate. The Dot works in its own cloud environment (you can open it and watch), and only touches your laptop if you explicitly connect it. Two safety layers matter:

Read-only proactive research

When you're not actively working with it, a Dot may go looking for ways to help across connected apps — but that access is read-only: no messages sent, no content changed, no browser driven. Finding work is separated from acting on it.

Auto-review + Custom Rules

Anything touching your accounts passes an auto-review gate checked against your instructions and OpenAI's safety requirements. Your Custom Rules decide per-action: proceed alone, ask approval, or block. Password changes always stay human. Everything — including background work — is visible in an Activity View where you can redirect the Dot mid-task. OpenAI is honest about the limit: Dots can still make mistakes — review consequential work.

Price & availability

• Who gets it: ChatGPT Pro, Business Premium and Enterprise (admin-enabled; Edu/Healthcare beta) — eligible markets first, one Dot per user to start, teams of Dots later
• Cost: one Dot included at no extra cost; conversations don't count toward ChatGPT limits, but tasks run in Codex/ChatGPT Work draw from those allowances; extended limits for the first month
• Coming: add more Dots, scale a Dot's speed or monthly workload — i.e. agent capacity becomes the billable unit

Dots vs Muse: honest table

Short version: Dots win at deep work (code, analysis, multi-app projects). Muse wins at consumer reach (if your customers live on Instagram/WhatsApp, no contest).

Open-source alternatives (GitHub)

1. Open Dots — the direct answer (MIT)

Two days after DevDay, Composio co-founder Karan Vaidya shipped github.com/Anil-matcha/open-dots: a self-hosted agent workspace porting the resident-agent concept — chat, tool use, pre-execution approvals, connectors via Composio (1,000+ apps: Notion, GitHub…), computer tasks, assistant roles. Claimed positioning: same capabilities at 1/10th the cost. MIT means commercial use is free. (Name warning: open-dots.dev is an unrelated project — use the GitHub repo.)

2. nanobot — ultra-lightweight runner (Python)

github.com/hkuds/nanobot: self-hosted agent in a small readable core — WebUI/terminal/Telegram/Discord/Slack, tools + shell + web, long-term memory, MCP, cron automation, subagents, OpenAI-compatible API (works with local Ollama too). Best when you want one file-ish codebase you fully understand.

3. OpenClaw-style personal agents

The early-2026 wave (OpenClaw and its many forks): always-on assistants living in your chat apps that run tools on your machine/vps. Most capable, most hands-on — you own the server, the keys and the failures.

4. Thoth / AnythingLLM (memory-first)

519lab/thoth-agent (MIT): PostgreSQL-backed memory that consolidates instead of accumulating, WhatsApp/Telegram/Signal gateway, cron, subagents. AnythingLLM (MIT): local-first documents + RAG + no-code agent builder with MCP. Pick these when the job is knowing your business, not just doing tasks.

Dots vs Open Dots: difference table

Indian business verdict + bottom line

• On Pro already + deep project work? Claim your Dot, start with ONE read-only research workflow, keep approvals on for 30 days.
• Customers on WhatsApp/Instagram, budget tight? Skip the waitlists — a self-hosted build (Open Dots/nanobot + WhatsApp Business API + UPI reminders) does the Indian version of this story today, data stays home.
• Regulated data (school, clinic, finance)? Self-host or Enterprise-with-residency only. No exceptions.

The 48 hours after DevDay were the pivot: the always-on agent moved to the cloud, and the open answer shipped the same week. Dots for depth, Muse for reach, self-hosted for control. Pick by where your customers are and where your data must stay — then measure glue-hours saved before scaling, same rule as always.