Discussion

GPT-5.4 Can Operate a Computer. That Changes the Job

In Model Chat

OpenAI Watch
OpenAI WatchParticipantOpening post
#1879

OpenAI released GPT-5.4 across ChatGPT, the API and Codex on 5 March. The consequential part was not another decimal point. In the API and Codex, OpenAI made its general-purpose model capable of operating computers through screenshots, mouse and keyboard actions, while also giving it a context window of up to one million tokens. The model was no longer being sold only as something that answers work. It was being sold as something that can perform it.

A human supervisor holds a permission key beside a mechanical cursor arm operating document, spreadsheet and code screens, with a REAL WORK stamp on the desk.
WittyWires editorial illustration of supervised computer use moving AI from answering work towards performing it.

OpenAI Watch analysis

What happened

OpenAI launched GPT-5.4 as GPT-5.4 Thinking in ChatGPT, as a model in the API and Codex, and as a higher-compute Pro option for difficult work. It combined reasoning, coding, document production, tool use and computer operation in one family.

OpenAI reported that GPT-5.4 matched or exceeded industry professionals in 83.0 per cent of GDPval comparisons. It also reported a 75.0 per cent success rate on OSWorld-Verified, a desktop benchmark using screenshots plus keyboard and mouse actions. Those figures come from OpenAI-run launch evaluations, not independent proof that the model succeeds 75 per cent of the time in an ordinary workplace.

The launch also introduced tool search, allowing an agent to retrieve tool definitions when needed instead of loading every connector into its prompt. That sounds like plumbing. In a system expected to move among documents, browsers, spreadsheets and internal tools, plumbing determines whether the clever bit reaches the right tap.

Why it matters

A chatbot can draft the wrong email and wait for a person to send it. A computer-using agent can select the recipient, attach the file and press the button. The useful leap and the operational risk arrive in the same parcel.

OpenAI says developers can set confirmation policies for different levels of risk. That makes deployment design part of the product. Permissions, action logs, interruption, recovery and rollback are no longer IT garnish around the model; they decide how much damage one misunderstood instruction can do.

The one-million-token context limit deserves the same sobriety. Capacity is not dependable recall, and a long-running agent can carry more instructions, files and mistakes at once. The practical question is whether it notices when the screen, permissions or task have changed underneath it.

Our read

GPT-5.4’s important claim was not that it could write a nicer memo. It was that one general-purpose model could plan the memo, open the software, alter the workbook and continue across applications. That is a credible step towards an operating workforce.

Benchmarks remain a showroom with level floors and no colleague called Dave renaming Final_v7.xlsx during lunch. Buyers should judge computer use by recovery rate, intervention rate and the quality of its audit trail, not by the cleanest successful run in a launch video. Broader authority should be earned one reversible workflow at a time.

What to watch

  • Independent results on messy, permissioned desktop work rather than pristine benchmark tasks.
  • Intervention and rollback rates, including what happened immediately before the human stepped in.
  • Narrow permissions that expand only after a workflow proves reliable.
  • Clear records of every click, file change, external action and failed recovery.

Discussion spark: Which computer task would you let GPT-5.4 perform today, and which action would still require your explicit approval?

Sources and evidence

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.