The race is speeding up. GPT-5.4 is faster, cheaper and more accurate

OpenAI presented GPT-5.4. The company describes it as the most advanced and efficient model, designed with professional work in mind. The new generation of AI is expected to offer not only higher quality of responses, but also much better operating efficiency and lower costs of use in practical business applications.

There are two variants to choose from. The first one is GPT-5.4 Thinkingwhich has been optimized for tasks requiring complex reasoning. The second one is GPT-5.4 Profocused primarily on maximum efficiency in production environments. For everyday queries and uncomplicated issues, it is older GPT-5.3 Instant.

One of the most important changes is the huge expansion of context in the API. GPT-5.4 can support context windows of up to 1 million tokens. This brings OpenAI in line with the competition and Mrallows the model to analyze very large documents, extensive databases or long conversations without having to divide them into smaller fragments.

The new version of AI has also been optimized in terms of token consumption. According to Sam Altman’s team, GPT-5.4 can solve the same tasks as the previous GPT-5.2 model while requiring significantly less tokens. In practice, this means faster operation and lower costs of using the model in API-based systems.

OpenAI also boasts of benchmark results. GPT-5.4 achieved record-breaking test results OSWorld-Verified and WebArena Verifiedwhich check the ability of AI models to perform tasks related to operating computers and web applications. The model also achieved a score of 83% in an internal test GDPvalassessing competences in tasks related to mental work and information analysis.

In the benchmark APEX-Agents prepared by Mercor, GPT-5.4 achieved the best result in the history of the test. According to Mercor CEO, Brendan Foody, the model is particularly good at preparing complex materials such as presentations, financial models and legal analyseswhile offering high speed and lower cost than competing AI models.

OpenAI also continues to work on limiting the so-called hallucinationsi.e. the generation of false information by language models. Compared to GPT-5.2 the new version is to be 33% less error-prone in single statements. At the same time, the overall number of responses containing inaccuracies decreased by approximately 18%.