OpenAI Launches GPT-5.4 with Pro and Thinking Versions

OpenAI has launched GPT-5.4, its most advanced and efficient model for professional work, available in Pro and Thinking versions.

OpenAI Launches GPT-5.4 with Pro and Thinking Versions
OpenAI logo with the GPT-5.4 icon

On March 5, 2026, OpenAI unveiled GPT-5.4, described as "our most capable and efficient frontier model for professional work". The model is available in two main versions, GPT-5.4 Pro and GPT-5.4 Thinking, each designed to meet specific needs in a professional context.

GPT-5.4 Versions: Pro and Thinking

The Pro version of GPT-5.4 is optimised for high performance, making it ideal for tasks that demand speed and precision. It is particularly suited to applications that need fast and accurate responses, such as processing complex data or generating content in real time.

The Thinking version, on the other hand, is designed to tackle tasks that require in-depth reasoning and advanced contextual understanding. It is ideal for activities involving complex analysis, such as drafting detailed reports, strategic planning or solving multi-step problems. This version lets the model process information more deliberately, improving the quality and consistency of responses in complex scenarios.

Technical Innovations and Performance

GPT-5.4 introduces significant technical improvements, including:

  • Extended context window: The API version of GPT-5.4 supports context windows of up to 1 million tokens, the largest OpenAI has ever offered, allowing longer and more complex texts to be processed without losing coherence.
  • Improved token efficiency: The model can solve similar problems using significantly fewer tokens than its predecessor, optimising resources and reducing operating costs.
  • Superior benchmark performance: GPT-5.4 achieved record scores on computer-use benchmarks such as OSWorld-Verified and WebArena Verified, as well as an 83% score on the GDPval test for knowledge-work tasks.

Professional Applications and Benchmarks

GPT-5.4 has demonstrated excellence across various professional areas, including:

  • Creating complex deliverables: The model excelled at generating presentations, financial models and legal analyses, delivering superior performance at lower cost than competing models.
  • Reduced errors: Compared with GPT-5.2, GPT-5.4 is 33% less likely to make errors in individual claims, and its overall responses contain 18% fewer errors, improving the reliability of the information provided.

Advanced Tool Handling and Safety

OpenAI has introduced significant improvements to tool handling and safety:

  • Tool handling via Tool Search: The new system lets models look up tool definitions as needed, reducing token usage and improving efficiency in API requests.
  • Chain-of-thought evaluation: A new evaluation has been implemented to test models' chain of thought, ensuring that the model's reasoning is transparent and reducing the risk of errors or deception in responses.

Conclusion

GPT-5.4 represents a significant advance in the evolution of language models, offering more powerful and reliable tools for professional applications. Its Pro and Thinking versions provide targeted solutions for different needs, while the technical innovations and improved benchmark performance make it an indispensable tool for professionals and businesses seeking efficiency and precision in their daily operations.

  • Extended context window: Supports context windows of up to 1 million tokens, improving the handling of complex texts.
  • Improved token efficiency: Solves problems with fewer tokens, optimising resources.
  • Superior benchmark performance: Achieves record scores on computer-use benchmarks and knowledge-work tests.
  • Advanced professional applications: Excels at generating complex deliverables such as presentations and legal analyses.
  • Advanced tool handling and safety: Introduces Tool Search and chain-of-thought evaluations to improve efficiency and transparency.

For more details on GPT-5.4 and its applications, you can consult OpenAI's official documentation.