OpenAI releases GPT-5.4: a bet on computer-using agents, not just code

On March 5, 2026, OpenAI introduced GPT-5.4, a new flagship that combines strong coding, reasoning and native computer control. The release shifts its emphasis from a "smart chatbot" to an AI agent that does more than answer questions: it can carry out chains of actions in interfaces and work tools on its own.

A bet on agents

The main news in GPT-5.4 is built-in Computer Use. The model works from screenshots, moves the cursor, clicks buttons, fills in forms and checks the result. For developers, this marks a more mature step from plain text generation to agentic scenarios: bots can browse websites, run browser steps and perform routine operations without a rigidly scripted sequence for every click.

This is no longer a feature for demo reels but a base layer for real business processes, where an agent needs to see the interface and confirm that an action actually worked.

According to OpenAI's official data, the gain is most pronounced where the model must act autonomously rather than just write code in a vacuum. On OSWorld-Verified, GPT-5.4 rose to 75.0% from 47.3% for GPT-5.2, and on BrowseComp to 82.7% from 65.8%. In SWE-Bench Pro, the improvement was more modest: 57.7% versus 55.6%. That is a good signal for those building assistants and workflow agents, and a more restrained one for those who expected an unqualified breakthrough in programming specifically.

• Native computer control via screenshots and UI actions
• Tool Search for large toolsets without bloating the prompt
• `xhigh` reasoning mode for heavy tasks
• Up to 1 million tokens of context in Codex, with experimental support
• A lower rate of factual errors than GPT-5.2

What the tests showed

In practice, the picture was less even than in the presentation benchmarks. In a visual test, the model was asked to build a complex smart-home dashboard, render the result itself and fix errors after a self-check. The overall composition and style looked recognizable, but the details let it down: text overlapped blocks, spacing drifted, some elements were clipped, and the promised neumorphic thermostat turned out to be a simplified circle. In other words, the model already catches the feel of an "expensive interface", but it is still far from an autonomous senior frontend developer.

The backend scenario was more convincing. Given a task to build a production-ready rate limiter for FastAPI and Redis, the model did not stop at a basic solution. It built a full design with strict typing, a Lua script for atomicity and a local fallback in case Redis goes down. In a logic test with conflicting scheduling constraints, it also got it right: instead of inventing "some" answer, it methodically proved that no solution exists. That is an important sign of maturity: the model more often recognizes a contradiction instead of confidently hallucinating.

Pricing and availability

OpenAI launched GPT-5.4 on March 5, 2026, simultaneously in the API, Codex and ChatGPT as GPT-5.4 Thinking. At release, the model began replacing GPT-5.2 Thinking for paying ChatGPT Plus, Team and Pro users, and GPT-5.4 Pro became available on Pro and Enterprise. For developers, this is arguably as important as the benchmarks: the new flagship did not stay a lab demo but went straight into working products.

API pricing looks more aggressive than you might expect from an OpenAI flagship: $2.50 per million input tokens, $0.25 per million cached input tokens and $15 per million output tokens. The model is more expensive per token than GPT-5.2, but OpenAI is betting on better token efficiency: if an agent solves a task in fewer steps and wanders less through the context, the overall economics can be quite reasonable even for small teams.

What it means

GPT-5.4 does not look like a "magic replacement for a developer", but it clearly shows where the market is heading. The next round of competition is not only about answer quality but about a model's ability to see an interface, use tools, remember a long context and carry a task through to a result. For business, the implication is simple: value is shifting from single prompts to agents that can work inside real processes. These are the scenarios that will now decide which platform is best suited for real work rather than one-off wow demos.