OpenAI Unveils GPT-5.4: Major Advances for Professionals

OpenAI Unveils GPT-5.4 with Pro and Thinking Versions

On Thursday, OpenAI announced the launch of GPT-5.4, a cutting-edge language model designed to deliver exceptional performance in professional settings. This model comes in several versions, including a standard version, a reasoning-focused version called GPT-5.4 Thinking, and a high-performance optimized version named GPT-5.4 Pro.

One of the standout features of GPT-5.4 is its ability to handle context windows of up to 1 million tokens. This advancement represents the largest context window ever offered by OpenAI, allowing for more efficient and seamless management of complex tasks that require a large amount of information.

Additionally, OpenAI highlighted the improved token efficiency in GPT-5.4. The model is capable of solving the same problems with a significantly reduced number of tokens compared to its predecessor, marking a notable gain in efficiency.

Enhanced Performance and Benchmarks

GPT-5.4 has demonstrated remarkable performance across various benchmarks, achieving record scores in the OSWorld-Verified and WebArena Verified assessments. These results underscore the model's ability to excel in complex and varied computational tasks.

The model also achieved an impressive score of 83% on OpenAI's GDPval test, which evaluates the cognitive capabilities of the model in intellectual work tasks. This score reflects the ongoing improvement in reasoning and information processing capabilities of GPT-5.4.

Moreover, GPT-5.4 topped Mercor's APEX-Agents benchmark, a test designed to assess professional skills in law and finance. According to Brendan Foody, CEO of Mercor, GPT-5.4 excels at creating long-term deliverables such as presentations, financial models, and legal analyses. He emphasized that the model delivers top-tier performance while being faster and less costly than competing leading models.

GPT-5.4 also continues OpenAI's efforts to reduce hallucinations and factual errors. The model is 33% less likely to make mistakes in individual assertions compared to GPT 5.2, and overall responses are 18% less likely to contain errors. These improvements are crucial for ensuring the reliability and accuracy of the information provided by the model.

Innovations in the API

As part of the GPT-5.4 launch, OpenAI introduced significant innovations in tool call management within the API. A new system, called Tool Search, has been implemented to enhance the efficiency of tool calls. Previously, system prompts had to define all available tool definitions when calling the model, a process that consumed many tokens as the number of tools increased. Now, the system allows models to search for tool definitions only when necessary, resulting in faster and less costly requests, especially in systems where many tools are available.

OpenAI has also integrated a new safety assessment to test the models' thought processes. This assessment aims to verify the continuous commentary provided by the models to demonstrate their reasoning process through multi-step tasks. AI security researchers have long expressed concerns about the possibility that reasoning models might misrepresent their thought processes. Tests show that this can occur under certain circumstances, but OpenAI's new assessment indicates that deception is less likely in the Thinking version of GPT-5.4. This suggests that the model does not have the capability to hide its reasoning, and monitoring the thought process remains an effective security tool.