OpenAI launches GPT‑5.4, pushing AI deeper into real work

OpenAI has launched GPT‑5.4, a new frontier model explicitly "designed for professional work" and rolling out across ChatGPT, the API and the Codex developer platform as of March 5, 2026. In ChatGPT, it appears as GPT‑5.4 Thinking, with a higher‑end GPT‑5.4 Pro variant for maximum performance on complex tasks, according to the company’s product announcement and model documentation. OpenAI is pitching the release as a consolidation of its latest advances in reasoning, coding and agentic workflows into a single work‑oriented system, a move that could reshape how knowledge workers interact with software day to day.

The model is also arriving with fresh competitive stakes. By targeting spreadsheet models, slide decks, long‑context research and code-heavy automation, GPT‑5.4 is aimed squarely at enterprise use cases where rivals like Google’s Gemini and Anthropic’s Claude have been jockeying for ground. As Axios reports, OpenAI is already wiring the new engine into tools that let ChatGPT work directly inside Excel and Google Sheets, signaling a bid to turn the chatbot into a full‑fledged work engine rather than a conversational sidekick.

What GPT‑5.4 actually changes

In its launch blog, OpenAI describes GPT‑5.4 as its "most capable and efficient frontier model for professional work," combining the reasoning capabilities of earlier GPT‑5‑series models with the industry‑leading coding performance of GPT‑5.3‑Codex in a single stack, while improving how the model uses tools and software environments for tasks like spreadsheets, presentations and documents. The company says GPT‑5.4 is its first general‑purpose model with native computer‑use capabilities, able to operate a computer via libraries like Playwright as well as by issuing mouse and keyboard commands in response to screenshots, with developers able to steer behavior via configuration messages and custom confirmation policies in high‑risk workflows, according to the official release from OpenAI.

On OSWorld‑Verified, a benchmark for navigating a desktop through screenshots and actions, GPT‑5.4 scores 75% task success, surpassing both GPT‑5.2’s 47.3% and reported human performance at 72.4%, the company says. OpenAI’s developer docs position GPT‑5.4 as its frontier model for "complex professional work," including long‑context analysis with up to a 1 million token window in some configurations, matching reports from third‑party observers such as DataCamp that highlight native computer use and tool discovery as headline features. An Axios report notes that early access is prioritized for developers and Codex users pursuing agentic workflows that automate tasks across web apps and internal systems.

Pricing details vary by tier, but early reviews suggest OpenAI nudged per‑token costs upward relative to GPT‑5.2 while arguing that overall task cost may fall because GPT‑5.4 can solve problems with fewer "reasoning" tokens, as analyzed by independent reviewers at Is It Good AI. Several community write‑ups and OpenAI’s own materials emphasize improved factual reliability, with one summary of the launch materials claiming the model is around one‑third less likely to make incorrect individual claims than GPT‑5.2—although those figures are based on internal evaluations that have not yet been fully detailed in peer‑reviewed work.

Implications for enterprises, workers and regulators

For enterprises, GPT‑5.4 is being framed as an "agent‑native" model that can operate across internal tools and public web services without bespoke glue code, potentially compressing entire workflows like month‑end reporting, due diligence memo drafting or multi‑system customer support into a single orchestrated prompt. OpenAI’s developer documentation stresses steerability and configurable safety behavior for these agents, allowing companies to dial in stricter confirmation requirements or tool‑access limits in finance, healthcare or other high‑risk domains, according to the API model page for GPT‑5.4.

The move deepens an existing trend: frontier AI models are increasingly evaluated on their ability to perform economically valuable knowledge work, with recent benchmarks like the AI Productivity Index (APEX) used to test whether systems can deliver gold‑standard outputs in business settings, as described in a working paper on high‑value knowledge work from the National Bureau of Economic Research. GPT‑5.4’s explicit focus on professional workflows further blurs the line between "assistant" and quasi‑autonomous colleague, especially when combined with native computer control.

Still, the very features that make GPT‑5.4 attractive to enterprises are likely to sharpen ongoing policy debates over safety and access. Agentic capabilities and native desktop or browser control increase the potential blast radius of model failures or misuse, even as OpenAI continues to stress configurable guardrails and internal safety testing in its launch materials. Regulators already scrutinizing "frontier" systems may see GPT‑5.4 as another data point in calls for capability‑based oversight, while competitors will face pressure to match or exceed its work‑specific performance without sacrificing reliability.

In practice, the release marks a shift from models that can be adapted to work tasks toward models that are explicitly engineered around them. For knowledge workers and developers, that may mean more leverage—and more pressure—to redesign their jobs around AI agents that no longer just answer questions, but log in, click, type and ship work on their behalf.