OpenAI has released GPT-5.4, a new foundation model promoted as its most capable and efficient for professional work. It features a massive 1 million token context window, improved token efficiency, and record scores on various professional and knowledge work benchmarks.
The model shows a significant reduction in factual errors and hallucinations compared to its predecessor. OpenAI also introduced a new Tool Search system for its API to make tool calling more efficient and conducted new safety evaluations on the model's chain-of-thought reasoning.
The main topics covered are the release and capabilities of GPT-5.4, its performance on benchmarks, improvements in accuracy and safety, and new API features.
On Thursday, OpenAI released GPT-5.4, a new foundation model billed as “our most capable and efficient frontier model for professional work.” In addition to the standard version, GPT-5.4 is also available as a reasoning model (GPT-5.4 Thinking) or optimized for high performance (GPT-5.4 Pro).
The API version of the model will be available with context windows as large as 1 million tokens, by far the largest context window available from OpenAI.
OpenAI also emphasized improved token efficiency, saying GPT-5.4 was able to solve the same problems with significantly fewer tokens than its predecessor.
The new model comes with significantly improved benchmark results, including record scores in computer use benchmarks OSWorld-Verified and WebArena Verified. The new model also scored a record 83% on OpenAI’s GDPval test for knowledge work tasks.
GPT-5.4 also took the lead on Mercor’s APEX-Agents benchmark, designed to test professional skills in law and finance, according to a statement from Mercor CEO Brendan Foody.
“[GPT-5.4] excels at creating long-horizon deliverables such as slide decks, financial models, and legal analysis,” Foody said in the statement, “delivering top performance while running faster and at a lower cost than competitive frontier models.”
GPT-5.4 continues the company’s efforts to limit hallucinations and factual errors. OpenAI said the new model was 33% less likely to make errors in individual claims when compared to GPT 5.2, and overall responses were 18% less likely to contain errors.
Disrupt 2026: The tech ecosystem, all in one room
Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to $400.
Save up to $300 or 30% to TechCrunch Founder Summit
1,000+ founders and investors come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately
Offer ends March 13.
As part of the launch, OpenAI has reworked how the API version of GPT-5.4 manages tool calling, introducing a new system called Tool Search. Previously, system prompts would lay out definitions for all available tools when calling the model — a process that could consume a lot of tokens as the number of available tools grew. The new system allows models to look up tool definitions as needed, resulting in faster and cheaper requests in systems with many available tools.
OpenAI has also included a new safety evaluation to test its models chain-of-thought, the running commentary given by the model to show its thought process through multi-step tasks. AI safety researchers have long worried that reasoning models could misrepresent their chain-of-thought, and testing shows it can happen under the right circumstances.
OpenAI’s new evaluation shows that deception is less likely to happen in the Thinking version of GPT-5.4, “suggesting that the model lacks the ability to hide its reasoning and that CoT monitoring remains an effective safety tool.”
Meridiano
