OpenAI Unveils GPT-5.4 With Pro and Thinking Versions

OpenAI has just released GPT-5.4, the latest and efficient frontier model for professional work and advanced reasoning. The model is available in three versions: a standard version, GPT-5.4 Thinking, and the premium GPT-5.4 Pro.

The company states that the latest launch is aimed at making AI systems highly accurate and powerful for current situations. On the social media platform X, OpenAI described GPT-5.4 as its most “factual and efficient” model to date, capable of delivering faster responses with fewer computational tokens.

The new model blends well the coding capabilities from the earlier GPT-5.3-Codex with improved support for spreadsheets, documents, and presentations.  This positions it squarely as a powerhouse for enterprise and knowledge work.

Moreover, the new model is 33% less likely to make errors in individual claims than GPT-5.2, and overall responses are 18% less likely to contain errors.  For an industry battling hallucination, those numbers matter.

GPT-5.4 Thinking emphasizes heavily on complex reasoning and multi-step problem solving. In ChatGPT, the model can outline its approach to complex tasks and allow users to modify instructions while it generates responses. This interactive approach allows users to guide the model during long analyses without restarting the process. 

The second variant, GPT-5.4 Pro, performs well for high-performance workloads that require maximum capability. It is built particularly for environments such as large-scale enterprise systems, software development workflows, and data-intensive projects. The Pro model prioritizes computational power and reliability for production environments. 

The GPT-5.4 delivers outstanding results on benchmarks.  GPT-5.4 scored a record 83% on OpenAI’s GDPval test for knowledge work tasks. That means the model matched or exceeded industry professionals in 83% of comparisons across 44 occupations. That’s a striking result.

Context capacity has also received a major update. The API version of the model will now support a context window of up to 1 million tokens, by far the largest available from OpenAI.  Developers building complex pipelines will find this especially useful.

Real-world partners are already impressed. Yashodha Bhavnani, Head of AI at Box, stated, “In Box’s evaluations comparing GPT-5.2 and GPT-5.4, we saw an overall performance improvement across our complex extraction tasks dataset from 72% to 78%. These gains demonstrate an increased ability to extract multiple pieces of information from a document in a single pass, including tasks requiring multi-step reasoning — a critical capability to drive and inform enterprise workflows.”

Third-party benchmarks require more validation. According to Mercor CEO Brendan Foody, “[GPT-5.4] excels at creating long-horizon deliverables such as slide decks, financial models, and legal analysis, delivering top performance while running faster and at a lower cost than competitive frontier models.”

The thinking version now includes a new feature that presents an upfront plan for its thinking. Hence, users can modify the response while it’s running. This prevents the model from going down the wrong path without burning through tokens.

OpenAI also addressed the major concern about transparency in reasoning. The company’s new safety evaluation found that deception is less likely in the Thinking version, “suggesting that the model lacks the ability to hide its reasoning and that CoT monitoring remains an effective safety tool.”

Industry observers believe that the launch is yet another step towards AI systems that can handle real knowledge rather than simple conversational tasks. The model works well for various tasks such as document analysis, financial modelling, coding, and long-form research across industries.

With the launch of GPT-5.4, APIs, and developer platforms, OpenAI works hard to integrate AI into professional workflows. The company’s emphasis on speed, reliability, and reasoning suggests that, in the future, there will be strong competition around AI models that are not only available for chat but also for real work.