OpenAI officially released the GPT-5.4 series model on March 5, 2026. Compared to previous versions, GPT-5.4 achieves a deep integration of reasoning capabilities, programming abilities, and agent workflows.

In the ChatGPT web version and App, the new Thinking (Reasoning) mode introduces a groundbreaking change in interaction:

Thinking Path Preview: Before giving a formal answer, AI will first display its "thinking plan." Users can see in real-time how the AI plans to break down the problem and can interfere and adjust the direction at any time during the generation process.

Deep Search Enhancement: For highly specialized and complex queries, the Thinking mode has stronger multi-turn web search capabilities, enabling it to cross-source compare and synthesize high-quality answers like human experts.

HCqs0sHaEAA0j8l.png

Productivity Engine: Inheriting Codex's Gene and Professional Scenario Optimization

GPT-5.4 officially merged the top programming capabilities of GPT-5.3-Codex, becoming OpenAI's first "mainstream reasoning model."

Office Software Collaboration: It has been specially optimized for scenarios such as spreadsheets (Excel), presentations (PPT), and various document processing, significantly reducing logical errors during the generation process.

Low Error Rate and High Factual Accuracy: Data shows that the factual error rate of GPT-5.4 is 18% lower than that of GPT-5.2, and false statements have decreased by 33%. Its performance in professional fields (such as finance and legal analysis) can now match that of human experts.

Agent Leap: Native "Computer Operation" and Million Context

GPT-5.4 is no longer just a chat window; it begins to possess true native computer operation capabilities:

Multi-Application Execution: The model can directly perform actions such as clicking, dragging, and inputting through screen shots and instruction writing in virtual environments or local software, completing complex workflows across applications.

Million Token Context: Supports an ultra-long context window of up to 1 million tokens (1M Tokens). This means the agent can "read" an entire large codebase or several quarters of financial reports at once, and then plan and verify long-term tasks lasting hours or even days based on this information.

Versions and Pricing

This update also launched the GPT-5.4Pro version, primarily targeting intensive professional developers and enterprise users:

Pro Version Exclusive Features: Supports higher intensity of reasoning effort (Reasoning Effort: High/X-High), focusing on extremely challenging scientific discoveries or architectural design tasks.

Availability