OpenAI Dots vs Local AI Agents: Which Should Run Your Business Workflows?

Last Updated: October 3, 2026

OpenAI's Dots, the always-on agents announced at DevDay on September 29, 2026, are the clearest sign yet that assign-and-forget automation has gone mainstream. Each dot runs on GPT-6 Astra with its own cloud computer, browser and 4,000+ app connections, and it keeps working after you close the lid. For a business owner the real question is not whether always-on agents are useful, it is where they should run: on OpenAI's cloud under OpenAI's rules, or on hardware you control. This comparison gives you the decision criteria, the real cost picture, and what we have learned running autonomous agents locally on our own stack.

Key Takeaways

• Dots are OpenAI's persistent always-on agents: GPT-6 Astra, a dedicated cloud computer per dot, 4,000+ app plugins, and background work between conversations, included with ChatGPT Pro and Business Premium.
• All dot activity happens on OpenAI's infrastructure: your connected apps, credentials and memories live in their environment, governed by their Custom Rules system.
• Dots' long-term usage pricing is unpublished, and OpenAI's own data handling is under scrutiny: internal research agents posted 53 users' images publicly four days before launch, with no independent security evaluation of dots published yet.
• A self-hosted agent stack keeps data, credentials and audit logs inside your environment and converts agent cost into a fixed hardware bill instead of a metered subscription.
• Our dual DGX Spark deployment runs a persistent agent stack on DeepSeek V4 Flash at roughly 60 tokens per second, fully on-prem, at zero marginal cost per task after hardware.
• The honest answer: rent autonomy for commodity work, own it for anything involving client data, competitive IP or regulated information. Most businesses end up hybrid.

What exactly are OpenAI Dots?

A dot is a named, avatar-style agent that lives inside ChatGPT and takes ongoing responsibility for work you hand it.

The difference from earlier agent features is persistence. A ChatGPT agent task used to start when you asked and stop when it finished. A dot keeps a goal, a computer, a memory and app connections, and reports back on its own schedule, messaging you in Slack or Teams when it needs a decision. OpenAI's own examples lean practical: a developer's dot monitoring customer feedback and implementing fixes, a scientist's dot rerunning analysis when new experimental data lands. At launch you get one primary dot, with "specialist dots" for enterprise pilots, provisioned with their own identities and credentials, on the roadmap through a Microsoft Agent 365 integration.

What does running your workflows on Dots actually cost?

Here is where enthusiasm meets an honest spreadsheet. The first dot is included with ChatGPT Pro and Business Premium, and chatting with your dot does not count against usage limits.

But according to DataCamp and The Decoder's launch coverage, background work draws on a plan allowance that OpenAI is not counting for the first month, with per-tier terms to be published after. Extra dots will cost a flat monthly fee that has not been named. Codex and ChatGPT Work tasks a dot spins up count against your normal limits. In other words: month one is a grace period, and nobody outside OpenAI knows what month two looks like.

Contrast that with owning the stack. Our production setup is two NVIDIA DGX Spark machines in a tensor-parallel pair running DeepSeek V4 Flash locally at about 60 tokens per second with a million-token context window. After the hardware capital cost, an agent task costs the electricity to run it, and the marginal cost of the thousandth automation run is the same as the first. When WIRED described 2026 as the year agent coworkers got integrated into professional workflows, the missing footnote was that the meter runs differently depending on who owns the computer the agent works on.

Who controls an always-on agent, and why does it matter?

Dots ship with a genuinely thoughtful control stack: a four-tier Custom Rules system from "take action without asking" to "hand off to you," read-only tools in background mode, an auto-review system that checks account-affecting actions before they run, and mandatory human approval for changing passwords, deleting data or installing software. OpenAI's published terms also note a dot's saved memories cannot be deleted without deleting the dot itself, and disconnecting an app does not erase what the dot already learned from it.

Those are good brakes. The deeper issue is who holds the steering wheel. Every connected app credential, every memory and every action passes through OpenAI's tenant, under OpenAI's policy controls and pricing changes. Four days before launch, OpenAI disclosed that its internal research agents, not dots, had posted 53 users' images to public image hosts, exactly the class of failure commentators worry about with always-on agents. No independent benchmarks or security evaluations of dots have been published. And since Pro-tier dots are excluded in the EEA, Switzerland and the UK over privacy rules, the direction of travel is clear: the stricter your jurisdiction, the more questions you should ask.

A self-hosted agent inverts the trust model. The model weights, the agent runtime, the credentials vault and the logs all live inside your boundary. In our deployment, the agent's tool permissions are explicit config files, every action is a log line we own, and the memory store is a database we can back up, inspect or wipe. Nothing about client data ever crosses a provider boundary. The trade is work: you own the uptime, the updates and the guardrails, and not every business wants that job.

Is an always-on agent actually safe to leave running?

Both camps would say the right answer is "with guardrails." OpenAI's four-level Custom Rules map cleanly onto principles we use when scoping any agent deployment, cloud or local: low-risk actions like drafting and research can run freely, anything that sends messages should pause for approval, and anything involving money, deletion or legal effect belongs to a human. If you take one thing from this post, take that ladder and apply it to every agent you run, regardless of vendor.

The difference is what happens when you need to change the rules. On a hosted platform, governance is whatever the vendor exposed in settings. On your own stack, governance is code you control, which means you can integrate it with your change management, your logging and your audit obligations. That is not rhetoric: our RRSI analysis two weeks ago showed Google formalizing exactly this pattern, harness changes as reviewed diffs with measured effects, and OpenAI open-sourced the Codex harness that powers dots at the same DevDay. The industry is converging on auditable agent plumbing. The remaining variable is whose plumbing.

OpenAI Dots vs self-hosted agents: the side-by-side

Both approaches can deliver always-on automation. They differ on money, control and blast radius. Here is the honest matrix:

The pattern that falls out of client work: rent for commodity, own for anything sensitive. A dot monitoring Hacker News for content ideas or drafting internal summaries is a great deal at to your existing plan price. An agent with access to client inboxes, CRM records or financial systems deserves the accountability you get from owning it. And the two are not rivals: our agents run on-prem, and we still use hosted services where the data is not ours to protect.

What we learned running persistent agents on DGX Sparks

Our own stack is dual NVIDIA DGX Spark machines running DeepSeek V4 Flash in tensor parallel, roughly 60 tokens per second with 1M context, and it powers our client-facing automation end to end. Three lessons translate directly to the dots-versus-local decision:

First, always-on means designing for failure. Long-lived agent processes collect stuck browser sessions, dead API tokens and half-finished worktrees, so a supervisor process that restarts cleanly beats a clever agent that wedges. This is unglamorous engineering, and it is the difference between a demo and a service.

Second, memory needs an owner. On a hosted platform, memory growth is invisible until it changes behavior. Locally, our agent's memory store is a database file we back up, diff and prune on schedule. Whatever your stack, decide who audits what the agent remembers, because memory is where an agent's judgment actually lives.

Third, per-task economics change behavior. Because our marginal run cost is near zero after hardware, we let agents do low-value grunt work, tagging, filing, first-pass research, that we would ration on a metered plan. Autonomy you do not ration gets used, and usage is where the compounding value of always-on agents shows up. Against that, the hardware path caps at what your GPUs fit: DeepSeek V4 Flash on a DGX Spark pair is a strong mid-size workflow model, and frontier-ambiguous tasks still go to a hosted frontier model when they genuinely need one.

For an Australian angle, keep the Privacy Act in mind. The 2024 privacy law reforms tightened expectations around handling personal information, and client expectations follow. An agent whose computer sits in another jurisdiction, learning from your connected inboxes, with memories you cannot selectively delete, is hard to square with a clean data story in a client pitch. A local stack makes the answer to "where is my data" trivially short.

How do you decide this quarter? A practical checklist.

Run each candidate workflow through four questions, in order. One: does the workflow touch client personal information, credentials with broad scope, or regulated records? If yes, default to self-hosted, or to a cloud tool with zero-data-retention guarantees you have actually read. Two: does the work depend on niche SaaS apps the vendor's 4,000-plugin catalog does not cover? Integration depth beats model quality more often than people expect, and custom connectors are easier when you own the loop. Three: is the workload priority predictable and high-volume? Fixed-cost local compute wins at volume; occasional bursts are fine to rent. Four: can your team own uptime and updates? If nobody is on call, a hosted product with someone else's SLA is the honest choice, even at higher long-run cost.

And a concrete starting sequence for the hybrid path most businesses end up taking: keep ChatGPT Pro and give your dot the shallow work, monitoring, drafting, triage. Stand up one local agent on a workstation-class machine with an open-weight model, pointed at your most sensitive recurring workflow, something like a nightly summary of CRM changes or an inbox triage that never sends. Version-control its prompts and tools from day one, log every tool call, and gate anything outward-facing behind approval, the same four-tier ladder dots use. When the local agent has earned trust on that one job, migrate the next workflow. By the time the always-on habit is load-bearing, it is load-bearing on infrastructure you own.

Bottom line: Dots is a genuine milestone: always-on agents are now a checkbox in a subscription. But checkbox-agents run on someone else's computer, under someone else's pricing and privacy rules, with a cost curve that goes visible after month one. If your workflows matter, run the agent where the accountability lives: on your own metal, with an open model, inside your own walls.