OpenAI introduces GPT-5.4 with more knowledge-work capability
New model offers 18% fewer factual errors and can control your keyboard and mouse via screenshots.
OpenAI has accelerated its release cadence with GPT-5.4, a model family explicitly designed for computer-use tasks and knowledge work. This launch comes at a critical moment as the company faces user defections to competitors like Anthropic, which saw record sign-ups following a public dispute over military AI use. The new models, including GPT-5.4 Thinking and GPT-5.4 Pro, are now available across ChatGPT, Codex, and the API, with Plus, Team, and Enterprise tiers gaining access to advanced reasoning features.
The technical upgrades are significant: a 1 million token context window matches top competitors, while improved token efficiency aids long-horizon tasks like web research. GPT-5.4 can now control a computer by issuing keyboard and mouse inputs based on periodic screenshots, a direct competitive move. It also features an 18% reduction in factual errors and enhanced visual understanding for images up to 10.24 million pixels. For developers, the API gains the Pro variant, and the 'Thinking' model offers users more upfront reasoning with the ability to change course mid-process. This release underscores OpenAI's strategy to compete on capability, cost, and efficiency to maintain its massive user base.
- Introduces computer-control capability, issuing keyboard/mouse inputs from desktop screenshots for agentic tasks
- Boasts an 18% reduction in factual errors and a 1 million token context window for long-context work
- Released amid user churn to Anthropic, which opened its memory feature to free users after a Pentagon deal controversy
Why It Matters
Directly enables AI agents for desktop automation and complex research, raising the stakes in the enterprise AI race.