ChatGPT, Codex, and the API now have access to OpenAI’s most recent frontier model, GPT-5.4. It is considered to have superior agents, thinking, and long-context professional work compared to its predecessor, as is the case with every new generation.
According to OpenAI, GPT-5.4 is a noteworthy achievement in which an AI model outperformed the human baseline in desktop navigation.
The regular GPT-5.4 and the more expensive GPT-5.4 Pro are the two variations of the product that are offered. According to OpenAI, GPT-5.4 eliminates the need for a separate code-specialised model in the majority of cases and can handle complicated knowledge tasks spanning more than forty professions, from engineering and data work to finance and law.
GPT-5.4 uses a mouse, keyboard, and screenshots to control desktop computers and web browsers. GPT-5.4 achieves 75% task success on the OSWorld-Verified benchmark for GUI navigation, surpassing GPT-5.2’s 47.3% and even a human baseline of 72.4%.
Additionally, the model demonstrated more dependable agents for website workflows by scoring higher on web navigation benchmarks like WebArena-Verified and Online-Mind2Web.
With a novel tool-search technique that reduces token usage in big tool ecosystems by almost half, OpenAI presents GPT-5.4 as its most token-efficient reasoning model to date. Approximately 1.05 million tokens can be processed simultaneously by the long-context version. It can process 272,000 input tokens in the ordinary option before additional fees for long context apply.
Similar to what Perplexity has been doing for some time, OpenAI’s system card for GPT-5.4 describes its internal thought processes. Additionally, OpenAI shows fewer false claims per response and lower hallucination rates than GPT-5.2.
The newly updated core competencies that users can get are
Native Computer Use: GPT-5.4 can actively carry out multi-step workflows through user interfaces, going beyond simple text recommendations.
One-Million-Token Context: The API version offers the biggest context window that OpenAI has ever provided, supporting up to one million tokens.
Agentic Tool Calling: Enhanced effectiveness in minimising latency during intricate operations like email management and spreadsheet processing by parallelising tool queries.
We use cookies to ensure that we give you the best experience on our website. If you continue to use this site we will assume that you are happy with it.
