OpenAI launches GPT-5.4, its most powerful AI model for professional work

OpenAI has launched GPT-5.4, available in ChatGPT as GPT-5.4 Thinking, as well as through its API and Codex, the company's coding platform. Alongside this release comes GPT-5.4 Pro, a variant aimed at delivering maximum performance on the most demanding tasks.

The new model brings together in a single version the coding capabilities previously offered by GPT-5.3-Codex, adding improvements in reasoning, external tool handling, and work with office documents such as spreadsheets and presentations. According to OpenAI's internal data, GPT-5.4 matches or outperforms human professionals in 83% of comparisons on GDPval, a benchmark evaluating specialised work across 44 different occupations, up from 71% for GPT-5.2. On spreadsheet modelling tasks typical of a junior investment banking analyst, the model scores 87.5%, compared to 68.4% for its predecessor.

One of the most notable new features is native computer control: GPT-5.4 can operate graphical interfaces, execute mouse and keyboard actions from screenshots, and work across different applications autonomously. In the OSWorld-Verified benchmark, which measures this capability, the model completes 75% of tasks successfully, surpassing human performance at 72.4% and GPT-5.2's 47.3%.

The model supports up to one million context tokens, allowing it to manage long workflows without losing coherence. It also introduces a feature called tool search, which lets the model locate and load only the tools it needs at any given moment, rather than processing all definitions from the start. In internal tests, this reduced token consumption by 47% without affecting accuracy.

On reliability, OpenAI reports that GPT-5.4 responses contain 33% fewer false statements and 18% fewer overall errors than GPT-5.2. In ChatGPT, the Thinking version now allows users to see the reasoning plan before the model generates its full response, with the option to redirect it mid-process.

GPT-5.4 Thinking is available for ChatGPT Plus, Team and Pro plans. Enterprise customers can enable it from the admin panel.

Key points

• OpenAI launches GPT-5.4, its most advanced model for professional use.
• Combines reasoning, coding and computer control in a single model.
• Matches or outperforms human professionals in 83% of evaluated specialised tasks.
• First OpenAI model with native computer control: mouse, keyboard and screenshots.
• Achieves 75% success rate on OSWorld, above human performance (72.4%).
• Supports up to one million context tokens for long workflows.
• Tool search reduces token consumption by 47%.
• Responses contain 33% fewer false statements than GPT-5.2.