OpenAI launches GPT-5.4, its most capable model yet
What's the story
OpenAI has announced the launch of its latest foundation model, GPT-5.4. The new model is being touted as "our most capable and efficient frontier model for professional work." In addition to the standard version, OpenAI is also offering a reasoning model (GPT-5.4 Thinking) and one optimized for high performance (GPT-5.4 Pro).
Enhanced capabilities
API version offers one million token context windows
The API version of GPT-5.4 comes with context windows as large as one million tokens, the largest ever from OpenAI. The company has also highlighted the model's improved token efficiency, saying it can solve the same problems with far fewer tokens than its predecessor. This makes it a more efficient tool for professional work, further enhancing its utility in various applications.
Performance milestones
New model aces several computer use benchmarks
The new model has also achieved record scores in computer use benchmarks OSWorld-Verified and WebArena Verified. It scored an impressive 83% on OpenAI's GDPval test for knowledge work tasks, setting a new standard for performance. The model also topped Mercor's APEX-Agents benchmark, which tests professional skills in law and finance, according to Mercor CEO Brendan Foody.
Professional applications
Model runs faster, costs less than rival frontier models
Foody praised GPT-5.4 for its ability to excel at creating long-horizon deliverables such as slide decks, financial models, and legal analysis. He added that the model delivers top performance while running faster and at a lower cost than competitive frontier models. This makes it an ideal choice for tasks requiring high levels of reasoning and efficiency in professional settings.
Error reduction
GPT-5.4 less likely to make errors
OpenAI has also focused on reducing hallucinations and factual errors with the new model. The company said GPT-5.4 is 33% less likely to make errors in individual claims compared to its predecessor, GPT-5.2, and overall responses are 18% less likely to contain errors. This improvement makes the model more reliable for users who depend on accurate information from AI systems.
New feature
Tool search reworks tool calling in API version of GPT-5.4
Along with the launch of GPT-5.4, OpenAI has also introduced a new system called Tool Search. The feature reworks how the API version of GPT-5.4 manages tool calling, allowing models to look up tool definitions as needed. This results in faster and cheaper requests in systems with many available tools, making it a more efficient option for users dealing with multiple tools at once.