Breaking: OpenAI has officially launched GPT-5.4, a landmark frontier model that introduces native "computer use" capabilities, effectively allowing the AI to take control of a mouse and keyboard to execute complex professional workflows. Released yesterday, March 5, the model represents a fundamental shift from chat-based assistants to fully autonomous agents capable of navigating desktop environments. With a massive 1-million-token context window and a new "Thinking" variant, GPT-5.4 is poised to redefine professional automation for developers and enterprises alike.
The Agentic Leap: Native Computer Use
The defining feature of GPT-5.4 is its ability to interact directly with software interfaces. Unlike previous models that relied on text-based API integrations, GPT-5.4 can "see" a computer screen through screenshots and execute actions such as clicking, typing, and scrolling. This capability allows it to handle end-to-end workflows—like researching data across multiple browser tabs, compiling findings into a spreadsheet, and generating a formatted presentation—without human intervention.
In internal testing, the model achieved a 75% success rate on the OSWorld-Verified benchmark, which evaluates an AI's ability to navigate operating systems. This score places it significantly ahead of competitors like Anthropic's Claude 3.5, which pioneered similar features last year. "This isn't just a chatbot anymore; it's a digital employee that sits at the desk with you," said an OpenAI spokesperson during the announcement. The feature is powered by enhanced visual perception capabilities, allowing the model to interpret dense, high-resolution user interfaces with pixel-perfect accuracy.
Powering Professional Workflows: 1 Million Tokens and 'Thinking' Mode
To support these complex, long-horizon tasks, OpenAI has expanded the context window to 1 million tokens. This massive capacity enables the model to ingest entire codebases, lengthy legal contracts, or extensive financial reports in a single prompt, maintaining coherence across hours of autonomous work. This addresses a critical bottleneck in previous "agentic" workflows, where models would often lose track of instructions during multi-step processes.
GPT-5.4 Thinking
Alongside the standard model, OpenAI introduced GPT-5.4 Thinking, available immediately to ChatGPT Plus, Team, and Pro subscribers. This variant replaces the previous "Thinking" model (formerly GPT-5.2 Thinking) and is optimized for deep reasoning tasks. It generates an upfront plan before executing actions, allowing users to audit the AI's logic before it commits to a task. This "chain-of-thought" transparency is crucial for enterprise adoption, where auditability is as important as performance.
For Developers: Tool Search and New API Pricing
For the developer community, the release brings significant efficiency upgrades. A new "Tool Search" feature allows the API to dynamically look up necessary tools rather than loading every possible function definition into the context window. OpenAI estimates this can reduce token usage by nearly 50% for complex agentic applications, drastically lowering operational costs.
Pricing for the standard GPT-5.4 API has been set at $2.50 per million input tokens and roughly $12 per million output tokens. While this is a premium over the previous generation's efficiency models, the cost is offset by the model's ability to complete tasks in fewer steps. Developers can also access GPT-5.4 Pro, a higher-tier model designed for maximum performance on the most demanding intricate tasks, though it comes at a higher price point.
Benchmarks and Safety Protocols
The performance metrics released by OpenAI paint a picture of a highly capable specialist. On the GDPval benchmark, which tests performance across 44 different professional occupations, GPT-5.4 scored 83%, an 8% improvement over its predecessor. It also showed marked improvements in coding, scoring 57.7% on the rigorous SWE-Bench Pro.
With great power comes great responsibility, and OpenAI has deployed a new "cyber safety stack" to monitor the model's autonomous actions. This includes asynchronous blocking and trusted access controls to prevent the agent from executing harmful commands or accessing unauthorized data. As the industry races toward fully autonomous AI, GPT-5.4 establishes a new baseline for what digital agents can safely achieve in the workplace.
