Claim. The product name is dots, lowercase. They are always-on ChatGPT agents on GPT-6 Astra, each with a cloud computer.
Method. chatgpt.com/features/dots/, read 2 October 2026.
As of.
OpenAI ChatGPT Dots vs Local Always-On Agents (Hardware + OSS)
By Billy G.R. · 2 October 2026
Key facts
- Checked
- 2026-10-02
- Author
- Billy G.R.
Checked 2026-10-02. Author: Billy G.R. Retail prices move; the hardware catalog stores the Amazon snapshot, not a promise of stock.
Citeable facts
Claim. Access on that page is Pro, Business Premium, and Enterprise, and Enterprise needs an admin. It is not described as a free feature.
Method. Same feature page. “Texting your dot is coming next” — messaging today is ChatGPT, Slack, and Teams.
As of.
Claim. A local stand-in at 27–30B Q4 needs about 18–20 GB before tool and browser overhead.
Method. 27 × 0.5 × 1.2 + 2 = 18.2. 30 × 0.5 × 1.2 + 2 = 20. Scaffold headroom of 1–4 GB is editorial, not part of the formula.
As of.
Methodology — parameter math, quantization bytes, and the source list.
ChatGPT dots are OpenAI’s always-on personal agents. The feature page spells the name in lowercase. They run on GPT-6 Astra. Each dot has its own cloud computer, starts from your ChatGPT memory, and uses connectors plus Custom Rules so you can allow an action, require approval, or block it. You message a dot in ChatGPT on web, mobile, and desktop, and the page also shows Slack and Microsoft Teams. Texting a dot is described as coming next, not as something you can do today. Docs: learn.chatgpt.com/docs/dots.
What the page is willing to claim
A dot keeps working between conversations, brings you work to review, and can be paused. Actions that touch accounts or share data go through Auto-review against your instructions, your Custom Rules, and safety checks. Changing a password stays with you. The page says you can connect your own computer, and that the dot’s own machine is a cloud computer. Access is Pro, Business Premium, and Enterprise in eligible markets. Enterprise users get it when a workspace admin turns it on. I am not calling it a free tier, and I am not quoting a dollar price from a page that sends you to plans elsewhere.
The examples on the page — decks, deal reviews, dinner orders — all assume OpenAI’s connectors and that cloud computer. A local stack does the subset you wire up: files you mounted, tools you wrote, a browser you launched. It does not inherit ChatGPT memory.
Always-on, without their computer
“Always-on” locally is a process supervisor, not a brand. I would run Ollama as a service, put Open WebUI in front of it, and let cron or systemd fire the jobs a dot would have nagged you about. Computer-use tasks go to OpenHands or browser-use. If you want several named agents in chat channels, CopilotKit OpenBot is the closer analogue than a single Open WebUI user. The model behind all of them should be one server, not a new copy per job.
A 7B model will answer. It will not reliably finish the multi-step chores the dots page demonstrates. Budget the 27–30B tier if the local agent is supposed to replace a dot for document work, and say so to yourself before you buy a 12 GB card.
Where the weights land
Same formula as the rest of the site. Q4_K_M is params × 0.6 + 2. The VRAM calculator adds a context slider the formula’s 1.2 factor does not fully cover. For a browser agent, add another 1–4 GB in your head for the scaffold. That add-on is my planning note, not an Apple or NVIDIA specification. The worked version is the agent VRAM guide.
| Unified / VRAM | What I would load | Formula |
|---|---|---|
| 8–12 GB | K2-Horizon-7B only | 6.2 GB. A 14B Q4 is 10.4 GB and leaves little for tools. |
| 24 GB | Qwen3.8-27B or Muse Glimmer Q4 | 18.2 GB or 20 GB. This is the practical “dot replacement” tier. |
| 36 GB | 27–32B with KV headroom | 32B Q4 is 21.2 GB. The extra memory is context, not a second model. |
| 128 GB | One large MoE, or a 32B plus a long browser session | Do not read 128 GB as “four dots.” It is one resident weight set unless you explicitly serve more than one model. |
A machine for that 24 GB tier
The Mac Studio M5 Max 36 GB (B0HGKSQMX6) was $2,449 on 1 October 2026 and is the box I would pick if the local agent is a daily driver. The EVO-X2 128 GB (B0F53MLYQ6, $3,649.99 the same day) is the “research bot and coding bot on one machine” purchase, with the bandwidth caveat in the Strix Halo guide: a lot of memory, not CUDA, about 256 GB/s class. A 32 GB RTX 5090 (B0DS2X13PH) showed $7,398 on 2 October 2026. It fits a 30B Q4. It does not fit a 70B Q4 at 44 GB.
Mac Studio M5 Max 36 GB on Amazon
ASIN B0HGKSQMX6. Amazon Associates link. The live price is on that page, not locked in here.
GMKtec EVO-X2 128 GB on Amazon
ASIN B0F53MLYQ6. Amazon Associates link. The live price is on that page, not locked in here.
Related guides
Sources
A correction or a hardware question goes to the contact page.