What is GPT-5.4, and what do its Thinking and Pro versions do?

Yesterday, Thursday, March 5, 2026, OpenAI released GPT-5.4, which it calls its most capable and efficient model for professional work. In ChatGPT it appears as GPT-5.4 Thinking, a reasoning model that takes time to work through a problem before replying. A stronger and more expensive GPT-5.4 Pro is available too. Developers can use both through the API and in Codex, OpenAI's coding agent.

Earlier this week, OpenAI updated its fast everyday model. This release sits at the other end of the range: the careful, slower models people choose for reports, analysis, code and multi-step jobs.

What OpenAI released

A token is a small piece of text, often part of a word. In the API, GPT-5.4 can take up to 1 million tokens of context, the text a model can hold in view at once; that is enough for several long books. TechCrunch notes it is by far the largest window OpenAI has offered. Inside ChatGPT, the window for GPT-5.4 Thinking stays the same as before.

OpenAI says the new model folds the coding strength of GPT-5.3-Codex, its specialist coding model from February, into a general model that also handles office work. The name change to 5.4 is meant to mark that merge.

How the new abilities work

A plan you can steer. For longer questions, GPT-5.4 Thinking now opens with a short outline of how it will tackle the task. If the plan is heading the wrong way, you can add instructions while it is still working, instead of waiting for a finished answer and starting over.

Computer use. This is the ability to operate software the way a person does: look at a screenshot, decide where to click, type, and check the result. OpenAI says GPT-5.4 is its first general-purpose model with this built in. On OSWorld-Verified, a test of completing tasks on a simulated desktop, it scores 75.0%, up from 47.3% for GPT-5.2 and above the 72.4% reported for people. Developers can also set rules for when the model must stop and ask for confirmation.

Tool search. Agents often connect to many tools, such as calendars, databases and file stores. Normally every tool's description is pasted into the prompt, which wastes tokens. With tool search, the model gets a short list and looks up a tool's full description only when it needs it. On one test with 36 tool servers connected, OpenAI says this cut token use by 47% with the same accuracy.

Office work and accuracy. OpenAI focused on spreadsheets, presentations and documents. On GDPval, a test of real work products across 44 occupations, GPT-5.4 matched or beat professionals in 83.0% of comparisons, against 70.9% for GPT-5.2. On prompts where users had flagged mistakes, its individual claims were 33% less likely to be false.

All these scores come from OpenAI, although some outside companies quoted at launch, such as Mercor, say it also leads their own tests of legal and finance work.

A worked example: a steerable research task

Because GPT-5.4 Thinking shows its plan first, it pays to write a request that invites correction:

I run a small bakery. Using the attached sales spreadsheet for 2025,
build a summary sheet showing monthly revenue, best and worst products,
and a chart of weekday versus weekend sales. Show me your plan before
you start, and flag any rows with missing or odd numbers instead of
guessing.

1. The model outlines its steps: clean the data, group by month and product, build the chart.
2. You notice it plans to treat holidays as weekdays and add, mid-response: "Count public holidays as weekend days."
3. It adjusts and continues, then lists the rows it could not use.

Why it matters

The release shows where OpenAI is placing its bets: on agents that finish real office and coding work, not only on conversation. Computer use, a huge context window and tool search all serve the same goal of letting a model carry a long task across many programs with less hand-holding.

It also raises the price of the top tier. GPT-5.4 costs more per token than GPT-5.2, though OpenAI argues it often needs fewer tokens for the same job. On safety, OpenAI treats GPT-5.4 as having High cybersecurity capability under its Preparedness Framework, its internal rulebook for serious risks, and deploys it with extra monitoring and access controls. It also published a new test suggesting the model is poor at deliberately hiding its reasoning, which helps researchers who watch that reasoning for signs of misbehavior.

Should you use GPT-5.4 Thinking or GPT-5.4 Pro?

For most paid ChatGPT users, GPT-5.4 Thinking is the right choice. It handles research, writing, spreadsheet analysis and coding well, shows a plan you can correct, and comes with Plus, Team and Pro plans.

GPT-5.4 Pro is for the hardest problems, where a better answer is worth a long wait and a high price: difficult research questions, complex legal or financial analysis, or tricky code. In the API it costs twelve times as much per token as the standard model, so developers should test whether the standard version is already good enough.

For quick everyday questions, neither is needed; the fast Instant model answers sooner. A sensible habit is to start with Instant, move to Thinking when a task has several steps, and save Pro for the rare job where the extra effort clearly pays off.