OpenAI launches GPT-5.4 with new reasoning and pro models

On Thursday, OpenAI released GPT-5.4, a new foundation model billed as its most capable and efficient model for professional work. In addition to the standard version, GPT-5.4 is available as a reasoning model, called GPT-5.4 Thinking, and an optimized high-performance version, called GPT-5.4 Pro. The company claims the model is significantly more reliable than its predecessor, GPT 5.2. According to OpenAI, GPT-5.4 is 33% less likely to make errors in individual claims compared to GPT 5.2, and overall responses are 18% less likely to contain errors.

For developers, the API version of GPT-5.4 introduces context windows as large as 1 million tokens, representing the largest context window available from OpenAI. It also features a new Tool Search system designed to manage tool calling. Previously, system prompts had to lay out definitions for all available tools when calling a model, which consumed a significant number of tokens as the tool library grew. The new Tool Search system allows the model to look up tool definitions only as needed, making the model more token efficient than its predecessor. This results in faster and cheaper requests in systems with many available tools.

The model has also shown strong performance on professional benchmarks, including record scores on the computer use benchmarks OSWorld-Verified and WebArena Verified. GPT-5.4 scored 83% on GDPval, which is OpenAI’s test for knowledge work tasks. It also took the lead on the APEX-Agents benchmark, a test provided by Mercor to evaluate professional skills in law and finance. Brendan Foody, the CEO of Mercor, stated that the model “[GPT-5.4] excels at creating long-horizon deliverables such as slide decks, financial models, and legal analysis.” Foody also noted that the model delivers top performance while running faster and at a lower cost than competitive frontier models.

OpenAI also addressed safety concerns regarding reasoning models. AI safety researchers have long worried that reasoning models could misrepresent their chain-of-thought—the running commentary showing the model’s thought process. However, OpenAI’s evaluations indicate that deception is less likely to happen in the Thinking version of GPT-5.4. According to OpenAI, this suggests that the model lacks the ability to hide its reasoning and that chain-of-thought monitoring remains an effective safety tool.

Why it matters

OpenAI’s latest release focuses on reliability and efficiency for professional tasks, signaling a shift toward models that can handle complex, long-horizon work with fewer hallucinations.