OpenAI has released GPT 5.4, a new foundation model that the company describes as its most capable and efficient frontier model designed for professional work.
Alongside the standard model, OpenAI is also offering two specialized versions: GPT 5.4 Thinking, which focuses on advanced reasoning tasks, and GPT 5.4 Pro, which is optimized for higher performance.
One of the biggest upgrades is the model’s context window. The API version of GPT 5.4 can handle up to 1 million tokens, making it the largest context window OpenAI has ever provided. This allows the model to process far larger documents, conversations, or datasets in a single request.
OpenAI also highlighted improvements in token efficiency. According to the company, GPT 5.4 can complete the same tasks using significantly fewer tokens than GPT 5.2. This makes it more efficient and potentially cheaper to run for developers and businesses.
The new model also achieved strong results on several performance benchmarks. GPT 5.4 recorded top scores in computer use tests such as OSWorld Verified and WebArena Verified. It also scored 83 percent on OpenAI’s GDPval benchmark, which measures performance on real-world knowledge work tasks.
In addition, GPT 5.4 performed strongly on Mercor’s APEX Agents benchmark, which tests professional-level skills in areas like law and finance. Mercor CEO Brendan Foody said the model is especially strong at producing complex deliverables that require long planning, such as slide presentations, financial models, and legal analysis. He also noted that GPT 5.4 delivers this performance faster and at a lower cost compared to other leading frontier models.
OpenAI has continued working to reduce hallucinations and factual mistakes in its models. The company said GPT 5.4 is 33 percent less likely to make errors in individual claims compared to GPT 5.2. Overall responses from the new model are also 18 percent less likely to contain mistakes.
The company also introduced a new feature called Tool Search for the API. In earlier systems, developers had to include the definitions of all tools directly in the system prompt when calling the model. This could consume many tokens when large numbers of tools were available. With Tool Search, the model can now look up tool definitions only when it needs them. This change can make requests faster and reduce costs in complex systems.
If this article helped you, please consider supporting our work. Every small contribution keeps Abijita.com independent and running.
OpenAI also added a new safety evaluation focused on chain of thought reasoning, which is the step-by-step explanation models sometimes provide when solving complex problems. Some AI researchers have raised concerns that reasoning models could hide or misrepresent their internal thinking. OpenAI’s tests suggest that this behavior is less likely in GPT 5.4 Thinking, indicating the model does not appear able to hide its reasoning and that monitoring the chain of thought remains a useful safety approach.
