A local model in OpenClaw vs a Dot on a frontier model: which jobs do you actually lose?

You lose the jobs that need a frontier model to keep going on a computer you are not tending: long browse-and-act work, a pile of connected apps, and a run that continues while your own machine is asleep. A local model in OpenClaw can still do a short file change you will read, and it can sit behind the chat apps you already use, on hardware you run. A Dot, as OpenAI described it on 29 September 2026, is the other product: an always-on agent on GPT-6 Astra with its own cloud computer. Pick the local OpenClaw path and you lose that unattended cloud job. Pick the Dot and you lose the “no hosted middle” path OpenClaw’s docs advertise. Those are different losses. They are not a single leaderboard.

I checked OpenClaw’s own docs and OpenAI’s announcement. Where a detail was not on the pages I could read, I say so instead of filling it in.

What each one actually is

OpenClaw, per docs.openclaw.ai, is a self-hosted gateway. You run it on your machine or a server you operate. It connects chat apps — the docs list Discord, Signal, Telegram, WhatsApp, Slack, iMessage, Matrix, Microsoft Teams, and others — to an agent. The docs say there is no hosted service in the middle. The model is a choice. The provider list includes local options: Ollama, llama.cpp, LM Studio, and vLLM, among many hosted APIs. The same docs tell you that for best quality and security you should use the strongest latest-generation model available. A local model in OpenClaw is allowed. It is not what the project recommends when you care about quality.

A Dot, per OpenAI’s 29 September 2026 announcement, is an always-on agent powered by GPT-6 Astra. OpenAI says it has its own cloud computer, can work toward your goals around the clock, learns from feedback over time, and can connect to apps through plugins — the post says over 4,000. The first Dot was included for Pro and Business Premium users in eligible markets. Enterprise, including Edu and Healthcare, was described as a beta an admin turns on. Press coverage of that launch also described messaging a Dot from ChatGPT and connecting Slack and Teams. I could not reload the full announcement page while writing this, so I am not going to invent quota numbers or the exact permission screens. The shape is enough: frontier model, cloud computer, always on, lots of app connections.

The jobs a local OpenClaw model loses

A long task that browses, decides, and keeps going. Local models drop the original request, invent tool names, and stop early. OpenClaw will not magically fix that because the weights are local. The gateway can route and schedule. The model still has to choose the tool and fill the arguments. If the job is “watch this site and finish the form,” a Dot on GPT-6 Astra is the setup OpenAI built for that, and a local model is the setup that will look finished before it is. You lose that job if you insist on local weights.

Work that must continue while your computer is off. A Dot’s cloud computer is the point of the announcement: it keeps working. A local model in OpenClaw runs where you put the weights and the gateway. If that is a laptop that sleeps, the job sleeps. You can put the gateway on a server you run, and you can even put a local model on that server if the hardware can hold it. That is still your host, your power bill, and your uptime. It is not the same as OpenAI keeping a cloud computer warm. If you will not run a server, you lose overnight work by choosing local OpenClaw.

The wide plugin catalog as a product. OpenAI’s post says a Dot can use a plugin ecosystem it numbers in the thousands. OpenClaw has plugins, channels, and a skills system. I am not going to claim a count I did not see on the OpenClaw docs I read, and I did not click through four thousand Dot plugins. If your job is “the integration for this SaaS already exists and I will not write it,” you might lose that job on a local OpenClaw box and have it on a Dot, or the reverse, depending on the app. Check the actual connector. Do not buy the number in the announcement as a promise about your app.

Unattended quality on messy instructions. OpenClaw’s docs already point you at the strongest model for quality. A local model will lose the jobs where being almost right is a bad edit: a permission change, a send, a booking. A Dot is a frontier model by OpenAI’s description. You lose those jobs in the sense that you should not assign them to the local model and walk away. You can still do them yourself, or point OpenClaw at a hosted frontier model, which is a third setup: OpenClaw’s gateway, not a local model. This article is local weights versus a Dot, not “OpenClaw can never call a strong API.”

The jobs you do not lose by staying local

A short edit in one directory, with list, read, and edit, and you reading the diff. A local model can be the only model for that. OpenClaw can host that agent if you want the chat-app front door. You do not need a Dot’s cloud computer to rename a key and look at the patch. Using a Dot for it spends an always-on frontier agent on a diff. You lose nothing important by refusing.

Chat from the channels OpenClaw documents, on a gateway you control. If the job is “message the assistant from Signal or Telegram and have the session stay on my hardware,” that is OpenClaw’s pitch, and a Dot is not obviously that product. I could confirm Slack and Teams in coverage of the Dot launch, and ChatGPT as the place you create the Dot. I could not confirm from the pages I read that a Dot sits in Signal or Telegram. If your job is those messengers, do not assume the Dot replaces the gateway. Read the current channel list before you migrate.

Keeping the accounts off a vendor cloud computer. OpenClaw’s docs say the gateway is yours and there is no hosted service in the middle. A Dot’s work, per OpenAI, happens on a cloud computer and can connect apps. Connecting those apps is the job you might want and the exposure you might not. For a research note on public pages, or a file edit, local OpenClaw does not lose the job. It refuses a different job: the one where the agent is signed into everything and keeps going.

The jobs you lose if you pick the Dot instead

You lose a loop that stays on hardware you run. The announcement is explicit about a cloud computer. Files, browser state, and connected apps in that design are in OpenAI’s setup, not only on your disk. If the job’s whole point was that the weights and the session do not leave, the Dot does not do that job. A local model in OpenClaw does, with the quality limits above.

You lose a quiet agent. A Dot is described as always on and working toward goals around the clock. That is a feature until the goal is vague and the apps are connected. I would not leave that running overnight on mail, calendar, or payments any more than I would leave any other cloud agent. If you needed an agent that does not continue while you sleep, the Dot’s design is the thing you lose by choosing it. You can perhaps constrain it. OpenAI’s launch materials, in secondary writeups, mention permissions and custom rules. I did not verify those screens, so I will not tell you they are sufficient. Assume an always-on agent stays capable until you have seen it stopped.

You lose the option to swap the model. OpenClaw’s provider list is the point of the gateway: local or hosted, your choice, and you can change it. A Dot, as announced, is GPT-6 Astra. If the job is “try this on a small model and only then call a strong one,” that job fits OpenClaw. It is not what the Dot announcement describes.

How I would assign real work

File edits I will read, on one directory, no shell: local model, OpenClaw only if I want a chat channel in front. I do not lose this job. I also do not need a Dot.

A gateway on a box I administer, with a frontier API when the tool call can change something, and a local model for the cheap reads: still OpenClaw, not a local-only setup, and not a Dot. I mention it so nobody hears “local OpenClaw loses long jobs” as “OpenClaw cannot call a strong model.” It can. The loss is specifically the local weights.

Unattended work across many apps, with someone else’s computer staying up: that is the Dot’s pitch. I would still withhold mail send, the password manager, and a calendar grant that can invite, the same as any other agent. The frontier model does not change the grant. If I cannot get a read-only connector, I do not connect it, and then I have lost that app on purpose.

Anything I cannot place in one of those rows does not get a default. “Which is better” is how people connect every app to the always-on agent and also expect the data to stay home. It will not do both. The jobs you lose are the ones that needed the property you did not pick.

What I would re-read before betting a workflow

OpenClaw’s docs change with the project. The provider list and the sentence about using the strongest model were on the docs I opened. If a later page contradicts the channel list, believe the later page. OpenAI’s Dot billing and the exact plugin you care about were not something I could fully reload. Before you move a real inbox or a company calendar, read the current consent screen. The announcement’s “4,000 apps” is not that screen.

Local OpenClaw loses unattended frontier work and any integration you do not actually have. It does not lose a reviewed file edit or a gateway you host. A Dot loses the no-hosted-middle path and a quiet, swappable model. Pick the loss you can stand. Do not pick both stories and hope the product you installed is the other one.