OpenAI Unveils New Model GPT-5.4 with Larger Context and Fewer Errors

Staff Reporter, PNN

As a new advancement in artificial intelligence technology, OpenAI has released its latest foundation model, GPT-5.4. The company claims that it is one of their most capable and efficient models so far for professional work.

The new model will be available in several versions. Alongside the general version, there is also a reasoning-focused “GPT-5.4 Thinking Model” and a high-performance optimized GPT-5.4 Pro version.

OpenAI stated that the API version of the model will support a context window of up to one million tokens, allowing analysis of very large documents, datasets, or long conversations simultaneously.

The company also claims that compared to previous models, GPT-5.4 can solve similar problems using significantly fewer tokens, which helps reduce both usage costs and processing time.

The model has also shown strong performance in various benchmark tests. It achieved record scores in computer-usage capability tests such as OSWorld-Verified and WebArena-Verified. In addition, it scored 83 percent on the GDPVal test, which evaluates knowledge-based professional tasks.

It also achieved top ranking in Marker’s Apex-Agents benchmark, which evaluates legal and economic analytical skills.

Brendan Foody, CEO of Marker, said the new model is particularly efficient in long-term professional tasks such as preparing presentation slides, financial models, or legal analyses. At the same time, it can complete tasks quickly and operate at comparatively lower cost.

OpenAI also stated that improvements have been made to reduce misinformation or “hallucinations.” According to their tests, GPT-5.4 is about 33 percent less likely to make errors in individual claims compared to the previous GPT-5.2 model, and the probability of overall incorrect responses is reduced by about 18 percent.

A new method called “tool discovery” has also been introduced in the API system. Previously, the model had to be provided with definitions of all tools at once, which consumed a large number of tokens. With the new system, the model can search for tool information as needed, making operations faster and cheaper.

As part of safety measures, OpenAI has also introduced a new evaluation method to verify the model’s chain-of-thought reasoning. According to the company, the thinking version of GPT-5.4 has a lower probability of providing misleading explanations, which could make safety monitoring more effective.