Back to all articles
•Technology

OpenAI Launches GPT-5.4, Featuring Enhanced Capabilities for Professional Work

View original source

OpenAI unveiled GPT-5.4 on Thursday, described as their most capable and efficient model for professional applications. The release includes a standard version, a reasoning model called GPT-5.4 Thinking, and a high-performance variant GPT-5.4 Pro.

  • The API version of the model now supports unprecedented context windows of up to 1 million tokens.
  • GPT-5.4 demonstrates improved token efficiency, able to solve problems with fewer tokens compared to earlier models.
  • It achieved record scores on benchmarks such as OSWorld-Verified, WebArena Verified, and scored 83% on OpenAI’s GDPval test for knowledge tasks.
  • Mercor’s APEX-Agents benchmark highlighted its superior professional skills in law and finance. According to Mercor CEO Brendan Foody, GPT-5.4 excels at delivering long-term deliverables like slide decks and financial models efficiently and cost-effectively.
  • The update reduces hallucinations and factual errors, showing a 33% decrease in individual claim errors and 18% better overall response accuracy compared to GPT 5.2.
  • OpenAI introduced a new Tool Search system, optimizing how tools are called in the API, saving time and cost by dynamically referencing tools as needed.
  • A new safety evaluation addresses AI reasoning transparency. The Thinking model version showed less tendency toward deceptive reasoning, reinforcing CoT monitoring as a reliable safety measure.
  • The article includes promotional content for TechCrunch Disrupt 2026 and TechCrunch Founder Summit 2026 opportunities, noting significant discounts for early registration.

The developments mark a significant stride in enhancing AI’s professional utility while addressing critical safety and efficiency aspects.