OpenAI launches GPT‑5.4 with integrated computer use and frontend UI capabilities.

OpenAI’s rollout of GPT‑5.4 marks the first time a single model combines the full suite of its recent advances in reasoning, coding, and agentic workflows, according to the company’s own announcement. The “Thinking” variant, now available in ChatGPT, the API, and Codex, adds a 1 million‑token context window and a persistent computer‑use (CUA) layer that can invoke external software without re‑prompting the model for each step. SQ Mah, an OpenAI researcher, explains that this CUA capability cuts token consumption by roughly two‑thirds on tasks that previously required repeated “think‑act‑observe” loops, because the model can retain state across multiple actions and only transmit the essential diffs back to the server. The same paper notes a boost in image understanding that lets GPT‑5.4 parse and manipulate website UI elements directly, enabling seamless generation of both visual mock‑ups and functional front‑end code.

The coding engine behind GPT‑5.4 builds on the GPT‑5.3‑Codex foundation, which was already touted as “industry‑leading” for software development. In practice, the new model can synthesize entire spreadsheet formulas, draft presentation decks, and edit complex documents with a single prompt, then verify the output by running the generated code in a sandboxed environment. OpenAI’s internal benchmarks, referenced in the release, show a 30 % reduction in average execution time for multi‑step programming tasks and a 25 % increase in correctness when the model is tasked with integrating third‑party APIs. By handling tool search and execution internally, GPT‑5.4 eliminates the need for developers to manually stitch together separate “search‑then‑code” pipelines, a workflow that has traditionally added latency and error‑prone hand‑offs.

For power users, OpenAI introduced a “GPT‑5.4 Pro” tier that unlocks the model’s maximum compute allocation. The Pro variant is positioned for “complex tasks” such as large‑scale data transformation, automated UI testing, and real‑time code refactoring across heterogeneous codebases. According to the same OpenAI blog post, Pro users see up to a 50 % speedup on tasks that involve iterating over thousands of rows in a spreadsheet or rendering high‑resolution UI components, thanks to the model’s ability to keep the CUA context alive across millions of tokens. This tier also grants higher‑resolution image processing, allowing the model to extract layout hierarchies from screenshots and generate corresponding HTML/CSS with pixel‑perfect fidelity.

The launch arrives amid a broader push by OpenAI to embed transactional capabilities into ChatGPT, as highlighted by recent CNET coverage of “instant checkout” integrations with platforms like Etsy and Shopify. While those features focus on e‑commerce, GPT‑5.4’s expanded tool use and UI comprehension lay the groundwork for more sophisticated, end‑to‑end workflows that could replace custom scripts and low‑code platforms. Wired’s reporting on OpenAI’s shopping add‑on underscores the company’s strategy to position ChatGPT as a universal interface for both consumer and enterprise tasks, a vision now technically feasible thanks to the model’s ability to manipulate software environments directly.

Despite the technical strides, the rollout has already sparked friction among existing users. VentureBeat notes that OpenAI recently withdrew several older models—including GPT‑4o and GPT‑3.5‑turbo—from the ChatGPT interface, prompting complaints from developers who relied on those versions for stability. The company’s leadership, including CEO Sam Altman, has framed the removal as a “necessary step” to focus resources on the newer frontier models, but the backlash highlights the risk of rapid iteration in a platform that powers millions of daily workflows. As OpenAI pushes GPT‑5.4 into production, its success will hinge on whether the promised efficiency gains translate into real‑world productivity for both hobbyist coders and enterprise teams.