GPT-5.4 Is Here: OpenAI’s ‘Most Capable and Efficient’ Model for Professional Work

OpenAI has released GPT-5.4, a new model designed to handle complex professional work more efficiently than earlier versions. The company says the model combines improvements in reasoning, coding, and automation, aiming to help users complete tasks faster and with fewer errors.

The company says the system is its “most capable and efficient frontier model for professional work,” marking a major upgrade to the capabilities available through ChatGPT, Codex, and its developer API.

The launch introduces two versions of the model: GPT-5.4 Thinking, focused on reasoning and step-by-step problem solving, and GPT-5.4 Pro, designed for maximum performance on demanding tasks.

Built for real-world knowledge work

OpenAI says the model performs particularly well in tasks that mirror everyday professional jobs, such as creating presentations, editing documents, and building spreadsheets.

In the company’s GDPval benchmark, which tests AI’s performance across 44 occupations, GPT-5.4 achieved strong results. The model matched or exceeded human professionals in 83% of comparisons, a significant improvement over earlier versions.

Performance gains also appeared in internal testing. For example, the model scored 87.3% on spreadsheet modeling tasks, compared with 68.4% for GPT-5.2. OpenAI also said human evaluators preferred presentations generated by GPT-5.4 68% of the time, citing stronger visuals and layout.

Native computer control introduces AI agents

One of the biggest changes in GPT-5.4 is its ability to interact directly with computers.

OpenAI says the model is the first general-purpose system it has released with built-in computer-use capabilities. This means the AI can perform tasks such as navigating applications, interacting with interfaces, and completing multi-step workflows across different programs.

The model can interpret screenshots, issue keyboard and mouse commands, and even write code to control software through tools like automation libraries. These capabilities point toward the growing trend of AI agents.

The company reported strong results on benchmarks measuring computer interaction. On OSWorld-Verified, a test designed to measure how well AI systems navigate desktop environments, GPT-5.4 achieved a 75% success rate, surpassing earlier versions.

Streamlining the developer experience

On the technical side, OpenAI is addressing the token tax that often makes complex AI agents expensive and slow.

A new feature called Tool Search allows the model to look up specific tool definitions only when it needs them, rather than loading every possible instruction into its memory at once. In one test across 36 servers, this approach reduced token usage by 47%.

Developers also get access to a massive 1-million-token context window in the API and Codex. This means the AI can remember and process much larger amounts of data at once, though OpenAI notes that prompts exceeding 272,000 tokens will be billed at twice the standard rate.

To help with the heavy lifting, a new “/fast” mode in Codex offers a 1.5x speed boost, helping developers stay in the flow while debugging or building apps.

Safety, factual accuracy, and pricing

OpenAI is also pushing back against the hallucination problem.

According to the company, GPT-5.4 is 33% less likely to make false individual claims compared to GPT-5.2. They’ve also introduced a new safety test to see if the model tries to hide its thought process. The results were encouraging for safety researchers, “suggesting that the model lacks the ability to hide its reasoning and that CoT monitoring remains an effective safety tool.”

The new model is rolling out now for Plus, Team, and Pro users. On the developer side, GPT-5.4 is priced at $2.50 per million input tokens for the standard model, with higher rates for output tokens and the more powerful Pro version. While the API pricing is higher than earlier models, OpenAI says improved efficiency and fewer retries could lower overall costs for many users.

If you’re still attached to the older GPT-5.2 Thinking model, you have until June 5, 2026, to use it before it is officially retired.

For more on the controversy surrounding OpenAI’s government partnerships, read our coverage of how ChatGPT briefly removed references to an OpenAI–Pentagon deal following public backlash.

The post GPT-5.4 Is Here: OpenAI’s ‘Most Capable and Efficient’ Model for Professional Work appeared first on eWEEK.