OpenAI launches GPT-5.4: New model hits 83% on pro-level knowledge benchmark

OpenAI GPT-5.4 launches with native computer-use and 1M tokens

interestingengineering.com

March 6, 2026

4 min read

🔥🔥🔥🔥🔥

46/100

Summary

OpenAI has launched GPT-5.4, a foundation model designed for professional and technical tasks, achieving 83% on a pro-level knowledge benchmark. The update features context windows of up to one million tokens and introduces native computer-use capabilities for AI agents, enhancing reasoning, coding, and workflow automation.

Key Takeaways

OpenAI launched GPT-5.4, achieving an 83% score on the GDPval benchmark for knowledge work tasks and posting record results on computer-use benchmarks OSWorld-Verified and WebArena Verified.
The model supports context windows of up to one million tokens and includes native computer-use capabilities, allowing AI agents to interact with software and execute complex workflows.
GPT-5.4 is reported to be 33% less likely to produce errors in individual claims compared to GPT-5.2, with overall responses being 18% less likely to contain factual mistakes.
OpenAI introduced a new system called Tool Search to reduce token usage and speed up responses by allowing the model to look up tool definitions only when needed.

Read original article

View story →