OpenAI Unveils GPT-5.4 with Pro and Thinking Editions
OpenAI has officially launched GPT-5.4, marking a significant advancement in AI technology. This new foundation model, announced on Thursday, is touted as the company’s most capable and efficient model suited for professional applications.
Versions of GPT-5.4
GPT-5.4 is available in three distinct formats:
- Standard Version
- Thinking Version – Optimized for reasoning tasks
- Pro Version – Designed for high performance
Advanced Features and Capabilities
The API version of GPT-5.4 boasts a remarkable context window of up to 1 million tokens, currently the largest offered by OpenAI. This feature promises enhanced efficiency in handling extensive data inputs.
Improvements in token efficiency are also noteworthy; GPT-5.4 achieves problem-solving capabilities with significantly fewer tokens compared to its predecessor, GPT-5.2.
Benchmark Performance
In testing benchmarks, GPT-5.4 delivered outstanding results:
- Record scores in OSWorld-Verified benchmarks
- High marks in WebArena Verified tests
- Record 83% on OpenAI’s GDPval test for knowledge work
- Leadership on Mercor’s APEX-Agents benchmark for professional skills in law and finance
Brendan Foody, CEO of Mercor, emphasized the model’s efficiency, stating it excels in generating long-term deliverables like slide decks and financial models while maintaining speed and cost-effectiveness against competitors.
Reduced Errors in Responses
One of the significant enhancements in GPT-5.4 is its ability to limit inaccuracies. The model shows a 33% reduction in individual claim errors compared to GPT-5.2, with overall response errors decreasing by 18%.
Tool Management Innovations
OpenAI has revamped the API’s tool management. The new Tool Search system allows users to call definitions as needed, a change that significantly decreases token consumption during model interactions. This adjustment aims to streamline requests, particularly when integrating multiple tools.
Safety Evaluations
In response to concerns about reasoning models, OpenAI has introduced a new safety evaluation method. This ensures that the model’s chain-of-thought remains transparent during complex tasks. The evaluation indicates that GPT-5.4’s Thinking version is less prone to deceptive practices, enhancing both safety and reliability in AI interactions.
The advancements with GPT-5.4 signify OpenAI’s commitment to developing AI that is not only powerful but also responsible and efficient. As the technology landscape evolves, these improvements present exciting opportunities for professional applications across various sectors.
