By Billy G.R. · 2 October 2026
Checked 2026-10-02. Author: Billy G.R. Retail prices move; the hardware catalog stores the Amazon snapshot, not a promise of stock.
Claim. Workspace Agents are OpenAI’s shared team agents. They are not personal dots.
Method. Announcement URL: openai.com/index/introducing-workspace-agents-in-chatgpt/. A re-fetch on 2 October 2026 timed out, so this page does not invent a connector or pricing list.
As of.
Claim. A team box should hold one shared 32B Q4 (about 21.2 GB) with room left, which is why 48–128 GB unified memory is the recommendation rather than one 24 GB card per seat.
Method. 32 × 0.5 × 1.2 + 2 = 21.2. Seats share the server. They do not each need a full second copy unless you choose to load one.
As of.
Claim. dots access, separately, is Pro, Business Premium, and Enterprise on the dots feature page.
Method. chatgpt.com/features/dots/, read the same day. Used here only to keep the products apart.
As of.
Methodology — parameter math, quantization bytes, and the source list.
OpenAI’s Workspace Agents announcement is the team product: shared agents for a Business, Enterprise, or Edu workspace, with workflows and connectors running in OpenAI’s cloud. Developer docs are indexed at developers.openai.com/workspace-agents. I tried to re-read the announcement on 2 October 2026 and the fetch timed out. I am not filling that gap with a feature list from secondary blogs. Stick to those two primary URLs for anything more specific than “shared, cloud, for a team.”
dots are personal always-on agents on GPT-6 Astra. Workspace Agents are the thing an admin shares across a company. Some coverage also says “ChatGPT Work.” I am not adopting that label. If a setting in your workspace uses a different name next month, the hardware advice on this page still applies: a team sharing one local model needs a bigger memory pool than a person running one dot replacement.
OpenAI’s version needs no GPU in the office. The moment the requirement is “the transcripts stay on our LAN,” you are shopping a server, not a ChatGPT seat.
The local picture I would actually deploy:
Five people do not mean five copies of a 32B model. They mean five KV caches against one resident weight set. The weights are the 21 GB. The caches are why a 24 GB card that “fits 32B” falls over when the second person pastes a long document.
| Seats | Shared model | Machine |
|---|---|---|
| One person, pilot | 7B Q4, 6.2 GB, or 27B Q4, 18.2 GB | 16–24 GB. Fine for a demo. Not a department. |
| A small team, shared 32B | 21.2 GB of weights | 48–64 GB unified so two or three long chats do not evict the model. |
| Team plus retrieval plus a bigger model | 32B or a 70B Q4 at 44 GB | 96–128 GB. Check the calculator before you promise 70B to the group. |
MoE models do not get cheaper because only some experts fire. DeepSeek-V4-Flash-0731 is about 284B total and 13B active. The active count is speed. The total is the disk and the RAM. A team server that “only uses 13B” and then OOMs is a planning bug, not a hardware failure.
For a team LAN host, the GMKtec EVO-X2 128 GB (ASIN B0F53MLYQ6) was $3,649.99 on 1 October 2026. Soldered LPDDR5X, no CUDA, bandwidth around 256 GB/s theoretical. It is the right memory size and the wrong expectation if someone on the team wants data-center tokens per second. I am not quoting a speed. A 24 GB single GPU is a pilot, then you outgrow it the week a second person pastes a spec.
GMKtec EVO-X2 128 GB on Amazon
ASIN B0F53MLYQ6. Amazon Associates link. The live price is on that page, not locked in here.
A correction or a hardware question goes to the contact page.