OpenAI has launched GPT-5.4, featuring advanced computer vision and improved performance across various tasks, enhancing user experience.
OpenAI has introduced GPT-5.4 and GPT-5.4 Pro, just days after the rollout of version 5.3 Instant. This latest version is now accessible via the ChatGPT interface, API, and Codex, with the GPT-5.4 Thinking variant available for Plus, Team, and Pro subscribers.
GPT-5.4 integrates advancements in reasoning, coding, and agentic workflows into a single model. pic.twitter.com/1hy6xXLAmJ
— OpenAI (@OpenAI) March 5, 2026
Pricing for the standard version starts at $2.50 per million input tokens and $15 per million output tokens, while the Pro version charges $30 and $180, respectively.
Improved Performance Metrics
In practical applications, GPT-5.4 demonstrates enhanced stability and quality. It scored 83% on the GDPval benchmark, which evaluates performance across 44 professions, significantly improving from GPT-5.2's score of 70.9%. Particularly notable is its performance in tasks typical of junior investment banking analysts, achieving 87.3% compared to GPT-5.2's 68.4%.
Evaluators favored presentations created by the new model 68% of the time due to superior aesthetics and effective image generation.
Moreover, GPT-5.4 is now the most accurate model from OpenAI, showing a 33% reduction in false individual statements and an 18% decrease in overall answer errors compared to its predecessor.
Introducing Computer Vision
This release marks the first inclusion of built-in computer vision capabilities and PC control features, enabling the model to interact with software based on screenshots and automate tasks through coding.
In the OSWorld-Verified benchmark, GPT-5.4 successfully completed 75% of tasks, outperforming GPT-5.2's 47.3% and human performance at 72.4%. Enhanced visual perception contributed to this success, with scores in understanding and logic tests rising to 81.2%, up from 79.5% in GPT-5.2.
Advancements in Programming
In the realm of coding, GPT-5.4 matches the specialized GPT-5.3-Codex but does so at a faster pace. The new Codex now features a /fast mode, accelerating generation by 1.5 times without compromising quality. Additionally, an experimental Playwright skill allows the model to visually debug web applications during coding.
Enhanced Tool Management
The introduction of a Tool Search feature streamlines plugin management by allowing the model to autonomously find and load necessary parameters, reducing token consumption by 47% without losing accuracy. Web search efficiency has also improved, with the Pro version achieving a benchmark score of 89.3%.
Improved Context Management
When faced with complex queries, GPT-5.4 Thinking presents a plan of action to users, facilitating real-time adjustments. This feature is currently available on the ChatGPT website and Android app, with iOS support expected soon. The model exhibits better context retention and coherence in long dialogues, enhancing its relevance and response quality.
Earlier this month, OpenAI faced user backlash over its contract with the Pentagon, highlighting ongoing concerns about the implications of AI technology.
