OpenAI launches GPT-5.4 with Pro and Thinking versions

On Thursday, OpenAI released a GPT-5.4, a caller instauration exemplary billed arsenic “our astir susceptible and businesslike frontier exemplary for nonrecreational work.” In summation to the modular version, GPT-5.4 is besides disposable arsenic a reasoning exemplary (GPT-5.4 Thinking) oregon optimized for precocious show (GPT-5.4 Pro).

The API mentation of the exemplary volition beryllium disposable with discourse windows arsenic ample arsenic 1 cardinal tokens, by acold the largest discourse model disposable from OpenAI.

OpenAI besides emphasized improved token efficiency, saying GPT-5.4 was capable to lick the aforesaid problems with importantly less tokens than its predecessor.

The caller exemplary comes with importantly improved benchmark results, including grounds scores successful machine usage benchmarks OSWorld-Verified and WebArena Verified. The caller exemplary besides scored a grounds 83 percent connected OpenAI’s GDPval trial for cognition enactment tasks.

GPT-5.4 besides took the pb connected Mercor’s APEX-Agents benchmark, designed to trial nonrecreational skills successful instrumentality and finance, according to a connection from Mercor CEO Brendan Foody.

“[GPT-5.4] excels astatine creating long-horizon deliverables specified arsenic descent decks, fiscal models, and ineligible analysis,” Foody said successful the statement,” delivering apical show portion moving faster and astatine a little outgo than competitory frontier models.”

GPT-5.4 continues the company’s efforts to bounds hallucinations and factual errors. OpenAI said the caller exemplary was 33% little apt to marque errors successful idiosyncratic claims erstwhile compared to GPT 5.2, and wide responses were 18% little apt to incorporate errors.

As portion of the launch, OpenAI has reworked however the API mentation of GPT-5.4 manages instrumentality calling, introducing a caller strategy called Tool Search. Previously, strategy prompts would laic retired definitions for each disposable tools erstwhile calling the exemplary — a process that could devour a batch of tokens arsenic the fig of disposable tools grew. The caller strategy allows models to look up instrumentality definitions arsenic needed, resulting successful faster and cheaper requests successful systems with galore disposable tools.

OpenAI has besides included a caller information evaluation to trial its models chain-of-thought, the moving commentary fixed by the exemplary to amusement its thought process done multi-step tasks. AI information researchers person agelong disquieted that reasoning models could misrepresent their chain-of-thought, and testing shows it tin hap nether the close circumstances.

OpenAI’s caller valuation shows that deception is little apt to hap successful the Thinking mentation of GPT-5.4, “suggesting that the exemplary lacks the quality to fell its reasoning and that CoT monitoring remains an effectual information tool.”