OpenAI's $500/Month Dots Agents Are Too Expensive — Here's How to Build Your Own AI Agent for Less

Yesterday, OpenAI announced Dots at DevDay — an always-on AI agent layer built into ChatGPT Pro. Persistent memory. Background tasks. Actions taken on your behalf while you're not looking. It sounds exactly like what people have been asking for.

The price tag: $500 a month.

That's $6,000 a year for an AI agent that lives on OpenAI's servers, runs on OpenAI's models, and disappears the moment you stop paying.

Here's what I want you to know before you pull out your card: you can learn how to build an AI agent that does most of the same things for somewhere between $15 and $40 a month — or nearly free if you have a spare Mac sitting around. I'm not a developer. I didn't write a line of code. Here's what that actually looks like.

---

What OpenAI Dots Actually Does — and What You're Paying For

Dots is OpenAI's answer to the question "can my AI agent keep working when I close the chat window?" The answer is yes — if you're on ChatGPT Pro at $500/month, introduced alongside the Dots launch.

What Dots gives you: agents that run in the background continuously, a persistent memory layer that accumulates context over time, and the ability to take actions — browsing, scheduling, sending messages — without you being in the loop. It's legitimately impressive engineering.

To put the price in context: $500/month is Netflix times roughly 30. It's a refurbished Mac Mini every six weeks. It's more than most people spend on their phone bill, car insurance, and a gym membership combined.

What you're paying for is convenience and zero setup. What you're trading away is flexibility, data ownership, and the ability to use any model you want. Your agents live in OpenAI's infrastructure, run on OpenAI's models, and your data — all of it — stays in OpenAI's system.

Worth noting: Anthropic launched Claude Cowork this same week — a directly competing feature that also offers persistent, always-on agents. It's included in Claude Pro at $20/month, or Claude Max at $100/month. Both products are doing roughly the same thing. One costs 5–25x less.

That gap is the whole story.

---

What a Self-Hosted AI Agent Actually Looks Like

Strip away the product branding and an AI agent is three things:

graph TD
    A[🧠 Model Brain\nDeepSeek V4 / Ollama / Claude API] --> B[⚙️ Agent Orchestration\nOpenClaw / LangGraph / CrewAI]
    B --> C[📋 Tasks & Outputs\nInbox monitoring, reports, schedules, web data]
    C --> B
    B --> D[💬 You\nSlack, email, voice]

The model brain is where the intelligence lives. This can be a cloud API call (cheap, no hardware) or a model running locally on your machine (free after hardware, requires a capable computer).

The orchestration layer is the system that tells the model what to do, when to do it, keeps track of memory, and handles the "always-on" scheduling. This is software — OpenClaw, LangGraph, CrewAI, or half a dozen other options depending on how technical you want to get.

Tasks and outputs are whatever you actually want the thing to do: monitor your inbox, draft summaries, pull data from the web, fire off a Slack message, generate a weekly report.

The "always-on" part — the thing that makes Dots sound magical — is not magic. It's just a computer that doesn't go to sleep. That's the entire secret. A Mac Mini you leave plugged in. A cheap VPS running in the cloud. Any machine that stays on runs your agent 24/7. Dots is charging you $500/month for someone else's always-on computer.

Tools worth knowing if you're looking at this seriously:

• Ollama — a one-line install that lets you run open-weight models locally. No command line expertise required beyond pasting one command. This is the tool that made local AI genuinely accessible to non-developers.
• LM Studio / Jan.ai — GUI apps for local model inference, even friendlier than Ollama for people who are truly command-line-averse.
• DeepSeek V4 — available both as a cheap cloud API and as open weights you can run through Ollama. More on this in the cost section.
• Mistral Large / Mistral Small — strong European alternatives with competitive pricing and solid performance on instruction-following tasks.

---

The Cost Breakdown: Dots vs. Building Your Own

Let's do the math properly.

Option A — Cloud API (No Extra Hardware Required)

Run your agent using API calls to a cheap, capable model. No new hardware needed if you already have a machine that can stay on.

Item	Monthly Cost
DeepSeek V4 API (heavy daily use)	~$8–15/mo
Hosting (existing machine or $5/mo VPS)	$0–5/mo
Agent orchestration software	$0–20/mo
Total	~$15–25/mo

Option B — Full Local (One-Time Hardware, Then Nearly Free)

Buy a Mac Mini once. Run everything locally. Pay essentially nothing after that.

Item	Cost
Mac Mini M2 (refurbished)	~$500 one-time
DeepSeek V4 local via Ollama	$0/mo
Electricity overhead	~$3–5/mo
Monthly cost after hardware	~$5/mo
Break-even vs. ChatGPT Pro at $500/mo	Month 1

This is not a trick. You spend $500 once on hardware. OpenAI Dots costs $500 in its first month. You break even immediately, and after that you're paying $5 a month in electricity while they're paying $500.

Option C — Hybrid (Best of Both)

Run lightweight, repetitive tasks locally on Ollama. Route complex reasoning to DeepSeek V4 API or Claude API only when you need it.

Item	Monthly Cost
Local inference (Ollama, routine tasks)	$0/mo
DeepSeek V4 API (complex tasks)	~$5–15/mo
Claude API (occasional use)	~$5–20/mo
Total	~$20–40/mo

graph LR
    A[Month 0] --> B["$500\nOpenAI Dots"]
    A --> C["$500\nMac Mini one-time"]
    B --> D["Month 1: $1,000 cumulative"]
    C --> E["Month 1: $505 cumulative"]
    D --> F["Month 12: $6,000 cumulative"]
    E --> G["Month 12: $560 cumulative"]
    style F fill:#ff4444,color:#fff
    style G fill:#22aa55,color:#fff

On DeepSeek V4 pricing: DeepSeek V4 is the current flagship from DeepSeek (V3 was retired in July 2026). On most practical tasks — summarization, drafting, research, planning — it performs at GPT-4o class while costing roughly 80–90% less per token. DeepSeek V4-Flash handles lighter, faster tasks at even lower cost. If you're building an agent that runs API calls throughout the day, the difference between GPT-4o pricing and DeepSeek V4 pricing is the difference between a $50 API bill and a $6 API bill.

Worth noting: using OpenAI's own API is dramatically cheaper than ChatGPT Pro for agent tasks. The $500/month subscription bundles unlimited usage in their interface — if your use case is API-driven agent automation, you're paying a massive premium for convenience.

---

The "Can I Actually Do This?" Section

Honest answer: it's not zero effort. But it's not coding, either.

Here's what setup actually involves:

1. Choose a machine — your existing Mac, a spare PC, or a ~$500 Mac Mini you buy once
2. Install Ollama — one command, takes five minutes, runs a model locally
3. Pick an orchestration tool — OpenClaw if you want something opinionated and documented, other options if you want more flexibility
4. Set up one or two tasks — tell the agent what to watch, what to do, what to output
5. Let it run

The real barriers are psychological, not technical. Are you comfortable installing an app? Are you willing to spend a Saturday afternoon on this? That's the actual bar.

What it is NOT:

• Writing Python
• Configuring servers
• Understanding APIs (optional, not required)
• Anything you'd need a developer friend to help with

The hardest part is deciding to try. The setup itself is easier than configuring your last router — and your router never did anything useful after you finished.

Ollama specifically deserves a mention here. Before Ollama, running a local model meant navigating PyTorch dependencies and CUDA drivers. Now it's a one-line install and a pull command. The access curve dropped dramatically, and it keeps dropping.

---

Where MyAIAgentOS Fits In

I've been documenting everything I learned building this kind of setup at MyAIAgentOS.com — it's a reference site for people who want to run their own agent stack without a developer background. The guides, the playbooks, the agent templates. Everything I figured out the hard way, written for people who are smart but not trying to become engineers.

If you want the full picture on open source AI agents and what a self-hosted setup actually looks like in practice, that's where I've been putting it.

---

Frequently Asked Questions

What is OpenAI Dots and how much does it cost?

OpenAI Dots is an always-on AI agent feature inside ChatGPT, launched at DevDay on September 29, 2026. It requires ChatGPT Pro at $500/month — the new pricing tier introduced alongside it. Dots lets agents run continuously in the background, take actions on your behalf, and maintain context over time. It is not available on cheaper ChatGPT tiers.

Can I build my own AI agent instead of paying for ChatGPT Dots?

Yes. A self-hosted AI agent using open-weight models like DeepSeek V4 or local inference via Ollama can replicate most of what Dots does for $15–40/month in API and infrastructure costs, or nearly free if you have a spare Mac or PC to run models locally. The setup requires a few hours but no coding experience.

How is Claude Cowork different from OpenAI Dots?

Claude Cowork (Anthropic's competing agent workspace, also launched this week) offers persistent agents that can take actions and run background tasks — directly comparable functionality to Dots. The key difference is pricing: Claude Cowork is included in Claude Pro at $20/month and Claude Max at $100/month. That's 5–25x cheaper than OpenAI's $500/month Dots tier for similar capabilities.

What is DeepSeek V4 and is it a good alternative to GPT-4o for running agents?

DeepSeek V4 is the latest flagship model from DeepSeek (released in 2026, succeeding V3 which was retired in July 2026). It delivers GPT-4o-class performance on most tasks — summarization, reasoning, code, instruction-following — at roughly 80–90% lower API cost. It's available as a cloud API and as open weights you can run locally with Ollama. DeepSeek V4-Flash is the lighter, faster variant for high-volume or low-complexity tasks.

What hardware do I need to run my own AI agent at home?

Almost any modern computer that can run 24/7 will work. A Mac Mini (M2 or later, refurbished around $500) is the most popular choice — quiet, power-efficient, and capable of running 7B–70B parameter models locally via Ollama. A $5/month cloud VPS also works if you prefer to use a remote API like DeepSeek rather than local inference. You do not need a GPU for most agent tasks.

Is building your own AI agent worth it compared to ChatGPT Pro?

It depends on what you value. If you want zero setup and don't mind the cost, ChatGPT Pro is the path of least resistance. If you want to own your data, customize your setup, use any model you choose, and cut monthly costs by 80–95%, building your own agent stack is increasingly accessible — even without coding experience. The break-even on hardware versus Dots subscription happens in month one.

---

Start Here

Want the step-by-step? I documented exactly how I set this up — no code required. Start here →

New guides drop regularly. If you'd rather get them in your inbox than remember to check back, the newsletter link is on the homepage.

---

Filed September 30, 2026.