OpenAI GPT-5.4: Native Computer Use, 1M Context, Benchmarks and API Pricing

OpenAI released GPT-5.4 on March 5, 2026, across ChatGPT (as GPT-5.4 Thinking), the API and Codex, alongside a higher-end GPT-5.4 Pro. OpenAI describes it as its most capable and efficient frontier model for professional work. The headline feature for developers is native computer use: GPT-5.4 is the first general-purpose OpenAI model that can operate a computer through screenshots and mouse and keyboard actions. This article summarises what changed, the published benchmarks, pricing, and what it means in practice, based on OpenAI’s announcement.

What is new

GPT-5.4 combines the coding abilities of GPT-5.3-Codex with improvements in reasoning, tool use and office-style knowledge work such as spreadsheets, presentations and documents. Key changes:

• Native computer use. In Codex and the API, the model can operate computers to carry out workflows across applications, either by writing code for libraries such as Playwright or by issuing mouse and keyboard commands in response to screenshots. Developers access this through an updated computer tool and can set custom confirmation policies for riskier actions.
• Up to 1M tokens of context in the API, so agents can plan and verify tasks over long horizons. In Codex the 1M window is experimental, and requests beyond the standard 272K window count against usage limits at twice the normal rate.
• Tool search. Instead of loading every tool definition into the prompt, the model receives a lightweight list and looks up definitions when needed. On 250 tasks from Scale’s MCP Atlas benchmark with 36 MCP servers, OpenAI reports this cut total token use by 47% at the same accuracy.
• Token efficiency. OpenAI says GPT-5.4 uses significantly fewer tokens than GPT-5.2 to solve problems.
• Steerability in ChatGPT. For longer queries, GPT-5.4 Thinking outlines its plan first, and users can add instructions mid-response without starting over.

Published benchmarks

OpenAI’s reported results, compared with GPT-5.2:

Benchmark	What it measures	GPT-5.4	GPT-5.2
OSWorld-Verified	Desktop navigation via screenshots and keyboard/mouse	75.0%	47.3%
GDPval (wins or ties)	Knowledge work across 44 occupations	83.0%	70.9%
SWE-Bench Pro (public)	Software engineering	57.7%	55.6%
BrowseComp	Persistent web research	82.7%	65.8%
Toolathlon	Multi-step real-world tool use	54.6%	46.3%

On OSWorld-Verified, OpenAI notes that the 75.0% result exceeds the 72.4% human performance reported by the benchmark’s authors. GPT-5.4 Pro reached 89.3% on BrowseComp. Most evaluations were run at the highest “xhigh” reasoning effort in a research environment, so everyday results in ChatGPT can differ.

Factuality

On a set of de-identified prompts where users had flagged factual errors, OpenAI reports that GPT-5.4’s individual claims were 33% less likely to be false, and full responses 18% less likely to contain any error, than GPT-5.2. That is a relative improvement on a specific test set, not a guarantee of accuracy; outputs still need checking.

Availability and pricing

• ChatGPT: GPT-5.4 Thinking rolled out to Plus, Team and Pro users, replacing GPT-5.2 Thinking, which remained under Legacy Models until its retirement on June 5, 2026. Enterprise and Edu admins could enable early access. GPT-5.4 Pro is available on Pro and Enterprise plans.
• API: gpt-5.4 and gpt-5.4-pro.

Model	Input (per 1M tokens)	Cached input	Output
gpt-5.2	$1.75	$0.175	$14
gpt-5.4	$2.50	$0.25	$15
gpt-5.4-pro	$30	—	$180

Batch and Flex processing cost half the standard rate, and Priority processing costs twice the standard rate. GPT-5.4 is more expensive per token than GPT-5.2, so whether it costs more per task depends on how much its token efficiency offsets the higher price.

Safety

OpenAI says it treats GPT-5.4 as having “High” cybersecurity capability under its Preparedness Framework and deploys it with corresponding safeguards, including monitoring, trusted access controls and request blocking for some higher-risk use. It acknowledges that classifiers may produce some false positives. Details are in the system card linked from the announcement.

What it means in practice

• For agent builders, native computer use and tool search are the main changes. Computer-use agents can automate software without APIs, but they should run in sandboxed environments with confirmation steps for actions such as payments, deletions or sending messages.
• For knowledge workers, the gains are in spreadsheets, presentations and research. OpenAI reports an internal score of 87.3% on junior investment-banking modelling tasks, versus 68.4% for GPT-5.2.
• For cost planning, test real workloads: compare total tokens and success rate rather than list prices alone.

Limitations

Benchmarks are vendor-reported and mostly run at maximum reasoning effort. Computer-use agents can still misclick, misread interfaces or take unintended actions, so human oversight remains important. Long-context accuracy also drops at the far end of the window: on OpenAI’s own MRCR v2 8-needle test, accuracy falls from around 97% at short lengths to 36.6% in the 512K–1M range.

Sources

• OpenAI: Introducing GPT-5.4