OpenAI releases GPT-5.4 with 1M-token context and Thinking build

OpenAI released GPT-5.4 today, introducing two distinct builds, GPT-5.4 Thinking and GPT-5.4 Pro, and expanding the model's ability to work with far larger inputs by providing a 1 million token context window. The company called the release "our most capable and efficient frontier model for professional work."

The Thinking build is positioned as a reasoning-oriented variant and is available now to ChatGPT customers on Plus, Team and Pro plans. OpenAI says the update widens tool-use as well, enabling the model to call external capabilities and handle longer, more complex workflows without losing earlier context. For professionals who rely on the model for drafting long legal briefs, technical specifications or multifile codebases, the jump to a 1 million token context effectively allows the model to hold entire books, multi-hour transcripts or large code repositories in a single session.

Technically, a 1 million token context window is a substantial increase over prior mainstream models and enables sustained chains of analysis over thousands of pages of text. That capability changes the practical boundaries of what generative AI can do in tasks that require continuous reference to extensive material, such as contract review, literature synthesis or comprehensive debugging of large software projects. It also reduces the need for repeated context injection and the attendant engineering workarounds developers have used to stitch together model outputs.

OpenAI framed the update as targeted at professional use cases where accuracy, efficiency and tool integration matter. The company highlights both reasoning improvements and cost-performance gains, positioning GPT-5.4 as a model meant to move routine high-skill workflows from human-only processes to human-plus-AI teams. For enterprises, the immediate business question will be how to integrate the expanded context window into document management, compliance and security systems without increasing exposure to sensitive data.

Larger context windows and deeper tool integration magnify both utility and risk. Persistent access to entire data sets in a single session increases the stakes for data governance and operational security. Organizations using GPT-5.4 will need stronger access controls, logging and review processes to prevent leakage of confidential information, and they will need to reassess existing compliance practices for regulated industries such as finance, health care and law.

The release also raises questions about model behavior over long documents. Longer context can reduce errors caused by context fragmentation, but it can also create new failure modes when the model must reconcile contradictory information embedded across thousands of tokens. Researchers and customers will watch for empirical evaluations of hallucination rates, factual accuracy and latency at the new scale.

Competitors and academic groups are likely to accelerate their own efforts to match larger context windows and task-specific reasoning builds. Policymakers and regulators will face renewed pressure to clarify oversight for capabilities that enable automated analysis at enterprise scale.

For users, the headline is immediate: GPT-5.4 expands what a single AI session can include and promises sharper reasoning on complex tasks. For institutions, the release demands rapid attention to integration, safeguards and evaluation before relying on the model for high-stakes decisions.