OpenAI Unveils GPT-5.4: The “Professional” Frontier with Native Computer Control and 1M Token Context

As generative artificial intelligence becomes increasingly entrenched within enterprise workflows, OpenAI has formally unveiled GPT-5.4—a novel foundational model heralded as being expressly engineered for professional endeavors—alongside its high-performance counterpart, GPT-5.4 Pro. This iteration transcends the mere pursuit of conversational fluidity with humans, directing its focus entirely toward code generation, rigorous data analysis, and the orchestration of agentic workflows.

GPT-5.4 stands not only as OpenAI’s inaugural general-purpose model endowed with “native computer-use capabilities,” but it also demonstrates overwhelming advancements in both spreadsheet manipulation and the generation of compelling presentations.

As pivotal partners, notably Microsoft, begin integrating models from rival entities, OpenAI has acutely recognized that to inextricably secure its foothold within the enterprise sector, its models must evolve beyond merely proffering suggestions to unequivocally executing and completing tasks. Among the most profound technological breakthroughs of GPT-5.4 is its distinction as OpenAI’s foremost general-purpose model equipped with state-of-the-art, native “computer-use capabilities.”

Historically, artificial intelligence was largely confined to generating code or sequential steps within a text interface; however, GPT-5.4 empowers AI agents to autonomously operate computers, seamlessly executing labyrinthine workflows across a multitude of applications. According to OpenAI’s empirical data, within the OSWorld-Verified benchmark—a rigorous test of desktop navigation proficiency—GPT-5.4 achieved an exemplary success rate of 75.0%. This not only eclipses the 47.3% achieved by its predecessor, GPT-5.2, but also surpasses the human baseline of 72.4%. Consequently, it demonstrates an unparalleled capacity to interpret screenshots and flawlessly execute precise mouse and keyboard commands. For professionals heavily reliant upon productivity software, GPT-5.4 heralds a profoundly palpable upgrade.

• Investment Bank-Caliber Spreadsheet Proficiency: In a rigorous internal evaluation simulating the intricate spreadsheet modeling tasks typical of a junior investment banking analyst, GPT-5.4 secured a formidable score of 87.3%, vastly outperforming the 68.4% attained by GPT-5.2.
• Aesthetically Superior Presentation Generation: In assessments pertaining to the creation of presentations, human evaluators demonstrated a 68.0% propensity to favor the outputs generated by GPT-5.4. This preference is attributed to its sophisticated aesthetic design, a broader spectrum of visual variations, and its highly efficacious utilization of image generation tools.
• Transparent “Thinking Mode”: Within the ChatGPT interface, the innovative GPT-5.4 Thinking mode proactively delineates its cognitive roadmap, empowering users to dynamically recalibrate the model’s trajectory mid-generation. This ensures the ultimate output aligns exquisitely with the user’s requirements, thereby significantly mitigating the need for iterative revisions.
• A Drastic Curtailment of Hallucinations: GPT-5.4 stands as OpenAI’s most factually rigorous model to date. When juxtaposed with its predecessor, the probability of encountering an erroneous solitary statement has diminished by 33%, while the likelihood of an entire response harboring inaccuracies has plummeted by 18%.

For developers, GPT-5.4 accommodates an unprecedented context window of up to one million tokens, affording AI agents the capacity to orchestrate, execute, and validate complex tasks over extraordinarily protracted temporal spans.

Equally noteworthy is the inauguration of the “Tool search” functionality within the API.

Historically, endowing a model with an array of utilities necessitated the encumbrance of the prompt with exhaustive tool definitions, precipitating an exorbitant consumption of tokens. Presently, GPT-5.4 dynamically queries requisite tool definitions—a paradigm shift that, during the rigorous MCP Atlas benchmark, successfully precipitated a monumental 47% reduction in aggregate token utilization, whilst steadfastly preserving the model’s precision.

Regarding pricing and availability, premium subscribers across the ChatGPT Plus, Team, and Pro tiers are granted immediate access to GPT-5.4 Thinking, effectively supplanting the antecedent GPT-5.2 Thinking model. Within the realm of API pricing, despite its vastly superior token efficiency, the unit cost has experienced an upward adjustment: GPT-5.4 commands $2.50 per million input tokens (an increase from GPT-5.2’s $1.75), while output tokens are priced at $15.00 per million.