#Where is GPT-5.4 Pro’s actual “computer use” on macOS?

1 messages · Page 1 of 1 (latest)

narrow dove
#

I’m confused by how GPT-5.4’s “computer use” capability is being presented to end users.

I’m a ChatGPT Pro subscriber on macOS, using the latest desktop app and the latest GPT-5.4 Pro model. But in practice, I still do not have any real local computer-control capability.

At most, ChatGPT can read content from certain supported apps such as a terminal, text editor, or code editor. In some cases it can reference or suggest edits. But it cannot actually control my Mac in a general way. It cannot type into arbitrary apps, click through the desktop, or perform real local computer actions the way “computer use” seems to imply.

So I want a clear answer:
1. What exactly does “computer use” mean for regular ChatGPT Pro users on macOS?
2. Is this capability only available through Agent Mode, Codex, or the API, rather than the normal ChatGPT desktop app?
3. Is there any version of ChatGPT for consumers that can actually operate the local computer in a practical Clawbot-like way?
4. If not, why is the messaging framed so broadly?

Right now, the product language makes it sound like GPT-5.4 can operate a computer directly, but the actual macOS experience feels much narrower — more like limited app-context reading than real computer control.

So what is the real availability status of this feature for Pro users on macOS?

boreal isleBOT
#

Hi @narrow dove. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:

#

You’re right that the current macOS experience is much narrower than “full computer control,” and the docs you’ve provided don’t describe anything like a general local “Clawbot” for consumers. Based on those sources, here’s what is actually available and how it fits together.

1. What “computer use” actually means (in the docs you shared)

In OpenAI’s developer docs, “computer use” refers to a developer-facing tool and model (computer-use-preview) that can drive a computer-like environment by issuing actions such as click(x,y), type(text), scroll, etc. This is only exposed via the Responses API, not via the standard Chat Completions API or consumer UI. [Computer use guide]

The loop looks like this:

  1. You call the computer-use-preview model with the computer_use_preview tool and environment details (browser, mac, windows, ubuntu).
  2. The model returns computer_call items (e.g., “click here”, “type this”).
  3. Your own code executes those actions on a real machine or browser.
  4. You send back screenshots / state as computer_call_output.
  5. Repeat until done. [Computer use integration]

So “computer use” in this sense is not a built‑in, turnkey local controller in the ChatGPT desktop app; it’s an API tool that developers wire up to their own automation environment.

2. What regular macOS ChatGPT actually does today (“Work with Apps”)

For end users on macOS, the feature that looks closest to “computer use” is Work with Apps on macOS, which:

#
  • Lets ChatGPT read content from supported apps (IDEs, terminals, Notes, etc.) and
  • In some IDEs, generate diffs and apply code edits to open files. [Work with Apps overview]

Key points from the docs:

  • It relies on macOS Accessibility to query content from compatible apps, or a VS Code extension for VS Code. [How it works]
  • For text and code editors, ChatGPT includes the full content of open editor panes in the foremost window (up to a truncation limit), and focuses on your selection if you highlight text. [Included content]
  • For terminals, it includes the last ~200 lines of open panes, again with focus on any selection. [Included content]
  • In IDEs, you can ask ChatGPT to edit open files directly; it generates a diff you can review and apply, or auto‑apply, and you can undo via the editor or the ChatGPT UI. [Work with Apps overview]

This is exactly the “limited app‑context reading + code edits” behavior you’re seeing. It does not describe arbitrary clicking around the OS, typing into any app, or general desktop automation.

3. Is this only in Agent Mode, Codex, or the API?

From the sources you provided:

  • Computer use (CUA):
    • Only documented as an API tool (computer_use_preview via Responses API). [Computer use guide]
    • Requires you (or a product like Codex) to implement the action loop and actually control a machine.
#

  • Codex CLI / Codex app:

    • Codex is a coding agent that can read, change, and run code in a selected directory on your machine, via local shell and file tools. [Codex CLI; Codex app]
    • It uses tools like local shell (now marked as outdated in favor of the newer shell tool) to run commands locally, but again this is a developer / power‑user tool, not the general ChatGPT desktop UI. [Local shell]

  • ChatGPT for macOS (consumer app):

    • Has Work with Apps (reading content + code edits) and other UX features (companion window, shortcuts, etc.). [Work with Apps overview; macOS release notes]
    • There is no mention in these docs of a general “computer use” controller wired into the macOS app that would click around arbitrary apps or the desktop.

So, from the documentation you’ve provided, the true “computer use” capability is currently an API / agent‑integration feature, and tools like Codex or custom agents can expose it in more automated ways. The regular ChatGPT macOS app exposes Work with Apps, which is narrower.

4. Is there any consumer ChatGPT that can fully operate the local computer?

In the sources you shared, there is no description of a consumer ChatGPT product that:

  • Can arbitrarily click around your macOS desktop,
  • Type into any app, or
  • Perform general OS‑level automation in the way you describe (“Clawbot‑like”).

The closest consumer‑facing automation described is:

#
  • Work with Apps on macOS: contextual reading + code edits in supported apps. [Work with Apps overview]
  • Codex CLI / app: local code + shell operations in a project directory, driven by an agent, but still scoped to code and commands, not arbitrary GUI control. [Codex CLI; Codex app]

The docs you provided do not state that GPT‑5.4 Pro in the macOS ChatGPT app can perform general GUI automation of the whole system.

5. Why does the messaging feel broader than the reality?

The knowledge sources you shared don’t include the exact marketing copy you’re reacting to, so I can’t quote or analyze that specific language. What they do show is:

  • A powerful developer‑side “computer use” tool that can, in principle, drive a computer environment via clicks and typing, but only when a developer wires it up through the API. [Computer use guide]
  • A consumer macOS feature set that is explicitly scoped to:
    • Reading content from supported apps via Accessibility / extensions, and
    • Editing code via diffs in IDEs. [Work with Apps overview]

There’s no explicit statement in these docs that “GPT‑5.4 Pro on macOS can fully control your computer.” The broad “computer‑using agent” narrative in the API docs is about what’s possible when integrated, not what’s currently shipped as a turnkey feature in the macOS ChatGPT app.

6. Practical bottom line for a Pro user on macOS

Based strictly on the provided documentation:

#

  • As a ChatGPT Pro user on macOS, in the standard desktop app you get:

    • Work with Apps: contextual reading of supported apps and code‑edit diffs in IDEs. [Work with Apps overview]
    • General ChatGPT features (models, data analysis, etc.), but not a general OS‑level controller.

  • The full “computer use” loop (clicking, typing, scrolling in arbitrary interfaces) is documented as an API feature (computer-use-preview + computer_use_preview tool) that developers or tools like Codex can harness, not something the macOS ChatGPT app exposes as a generic “control my Mac” mode. [Computer use guide]

So your impression is accurate: the current macOS Pro experience is “limited app‑context reading and code edits,” not full local computer control, and the docs you’ve provided don’t show a broader, consumer‑ready local controller beyond that.

Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:

#

@narrow dove Kapa.ai is still learning and improving, please let me know how I did by reacting below.