OpenAI launches GPT-5.4 with two tiers and native computer-use

What Happened

OpenAI launched GPT-5.4 on March 5, 2026, in two distinct variants. GPT-5.4 Thinking is available to all paid ChatGPT subscribers — including the $20-per-month Plus plan — and replaces GPT-5.2 Thinking, which will be retired in three months. GPT-5.4 Pro is restricted to the $200-per-month Pro tier and Enterprise plan users, according to TechCrunch.

The release carries two headline improvements. First, GPT-5.4 is OpenAI's first mainline model with built-in computer-use capabilities: it can interact directly with software to complete tasks in what VentureBeat describes as a "build-run-verify-fix loop," meaning the model runs code, checks the output, and corrects errors without waiting for a human to intervene. Second, the API version supports a context window of 1 million tokens — by far OpenAI's largest ever — enabling analysis of entire codebases or long document collections in a single request, according to eWeek.

On accuracy, The New Stack reports that GPT-5.4 reduces individual factual errors by 33% compared to GPT-5.2, and makes error-containing responses 18% less likely overall. Additional improvements cover coding, multimodal tasks, complex multi-step workflows, and financial plugins for Microsoft Excel and Google Sheets, according to VentureBeat. The model is rolling out gradually to ChatGPT and is also available in the API and Codex.

Why It Matters

Until now, AI assistants like ChatGPT answered questions and generated text — but operating software on your behalf required specialized tools or developer-built workarounds. GPT-5.4's native computer-use capability changes that baseline. The model can, for example, open a spreadsheet, run a script against it, check whether the output is correct, and fix the code if it is not — a workflow previously requiring human intervention at each step. Anthropic introduced a comparable computer-use feature for Claude, and Google has equivalent agentic tools; GPT-5.4 brings that capability into OpenAI's mainline product for the first time.

The 1 million token context window is also significant in competitive terms. OpenAI's previous models topped out well below Gemini's long-running 1 million token ceiling — a frequently cited disadvantage. GPT-5.4 closes that gap. As a point of scale: one million tokens is roughly the equivalent of ten full-length novels processed in a single conversation.

The tiered release deserves careful attention. The gap between what a $20-per-month subscriber receives and what a $200-per-month subscriber receives has been widening with each OpenAI model cycle, and GPT-5.4 continues that pattern. For Plus users, GPT-5.4 Thinking is a meaningful upgrade delivered automatically. For access to the full Pro model, the cost remains ten times higher. It is worth noting that GPT-5.4 is an incremental release rather than a generational leap — the computer-use capability and error-reduction improvements are genuine, but they consolidate features that previously existed in specialized tools, according to benchmark observers at lmcouncil.ai.

Labor market research released this same week adds broader context for why model releases like this one matter beyond benchmarks. A Harvard Business School study tracking nearly all U.S. job postings from 2019 through March 2025 found a 13% decline in postings for routine automation-prone roles alongside a 20% increase in demand for analytical, technical, and creative positions since ChatGPT's launch, according to Harvard Business Review. Separately, Anthropic's Economic Index found that 49% of jobs now involve AI in at least a quarter of tasks, up from 36% in early 2025. The picture that emerges is not mass displacement but gradual reshaping — and GPT-5.4's expanded agentic capabilities are the kind of development that accelerates that shift.