OpenAI Launches GPT-5.4, Its First Model That Can Operate Your Computer
OpenAI released GPT-5.4 on Wednesday, a model the company calls its most capable system to date — and the first in its lineup with native computer use capabilities. The model can autonomously navigate desktops, browsers, and software applications, marking OpenAI's most aggressive entry into the agentic AI race that has consumed the industry over the past year.
The release consolidates capabilities that OpenAI had previously scattered across separate models. GPT-5.4 merges the coding strengths of GPT-5.3-Codex with improved reasoning and, crucially, the ability to take control of a computer to complete multi-step tasks across different applications without developer-built infrastructure. It can write code to operate software, issue keyboard and mouse commands in response to screenshots, and call upon external tools and APIs to execute complex workflows.
The numbers tell the story of incremental but meaningful progress. On GDPval, a benchmark testing professional knowledge work across 44 occupations, GPT-5.4 scored 83 percent — edging past Anthropic's Claude Opus 4.6, which achieved 78 percent. In SWE-Bench Pro, which evaluates real-world coding tasks, the model hit 57.7 percent. On the OSWorld-Verified benchmark for computer use, it reached 75 percent accuracy. OpenAI also claims a 33 percent reduction in false claims compared to GPT-5.2, with full responses 18 percent less likely to contain errors.
Alongside the model, OpenAI is rolling out ChatGPT for Excel and Google Sheets in beta — an embedded version of ChatGPT designed for building and analyzing complex financial models directly inside spreadsheets. New integrations with FactSet, MSCI, Third Bridge, and Moody's aim to let enterprise teams pull market and company data into a single AI-powered workflow.
That enterprise push puts OpenAI on a direct collision course with Anthropic, which launched its Claude for Financial Services suite last July and has been steadily expanding its foothold in corporate environments. Both companies are now racing to capture the same prize: becoming the default AI layer for professional knowledge work. The move could also rattle traditional financial data providers, whose stocks have already been battered by fears that AI tools will render legacy enterprise software obsolete.
"Developers don't just need a model that writes code. They need one that thinks through problems the way they do," said Mario Rodriguez, GitHub's Chief Product Officer, pointing to GPT-5.4's performance in multi-step, tool-dependent workflows. GitHub Copilot began rolling out GPT-5.4 integration the same day.
The model supports context windows up to one million tokens in the Codex app and comes in three tiers: GPT-5.4 for Plus, Team, and Pro subscribers; GPT-5.4 Thinking for reasoning-heavy tasks; and GPT-5.4 Pro for maximum performance, available to Enterprise and Edu users. API pricing lands at $2.50 per million input tokens and $15 per million output tokens. OpenAI says the model is more token-efficient than its predecessors, meaning it requires fewer tokens to solve many tasks despite the slightly higher per-token cost.
Perhaps the most telling feature is one aimed at user experience rather than raw capability. GPT-5.4 Thinking now provides an outline of its work for complex queries and lets users adjust their request mid-response — steering the model toward the desired outcome without starting over. It is a small but revealing acknowledgment that the real challenge in agentic AI is not just getting the model to act, but getting it to act the way the user intended.
The release arrives at a pivotal moment for OpenAI. The company, which recently secured a $100 billion funding round and outlined plans for $600 billion in total compute spending, needs GPT-5.4 to demonstrate that its massive financial bets are translating into products that businesses will actually pay for. With Anthropic, Google, and a growing field of agentic competitors circling the same enterprise market, the window for establishing dominance is narrowing fast.
GPT-5.4 is rolling out now across ChatGPT, Codex, and the API.
