OpenAI has released GPT-5.4, a new model that arrives just two days after the company launched GPT-5.3 Instant, introducing native computer control, spreadsheet integrations, and a 47% reduction in token usage on certain tasks compared to prior versions.
The model ships in two tiers: GPT-5.4 Thinking and GPT-5.4 Pro. Thinking is available to all paid ChatGPT subscribers, starting with the Plus plan at $20 per month. Pro is reserved for ChatGPT Pro ($200 monthly) and Enterprise users. Free-tier users can access the model only when their queries are automatically routed to it. Both versions are available in OpenAI’s API and its Codex development environment.
Native Computer Use
The release’s most consequential capability is native computer use, which OpenAI describes as the first time it has shipped this feature in a general-purpose model. Through the API and Codex, GPT-5.4 can operate a user’s computer by writing code via libraries like Playwright or by issuing direct mouse and keyboard commands in response to screenshots. The model is designed to carry out multi-step workflows across applications without human handholding at each step.
Benchmark numbers back the claim. On OSWorld-Verified, which measures desktop navigation, GPT-5.4 scored 75.0% success compared to 47.3% for GPT-5.2. Reported human performance on the same benchmark sits at 72.4%, meaning the model now exceeds it. On BrowseComp, GPT-5.4 improved by 17 percentage points absolute over GPT-5.2, with GPT-5.4 Pro reaching 89.3%. On Online-Mind2Web, the model reached 92.8% success using screenshot-based observations alone.
Vision and document handling also improved. GPT-5.4 scored 81.2% on MMMU-Pro without tool use, up from 79.5% for GPT-5.2, and achieved a 0.109 average error on OmniDocBench, down from 0.140. The model also supports image inputs up to 10.24 megapixels at an “original” detail level.
Spreadsheet Integrations
OpenAI is simultaneously releasing a suite of integrations that plug GPT-5.4 directly into Microsoft Excel and Google Sheets, enabling automated analysis and task completion at the cell level. The move positions the model inside the daily workflows of enterprise users, placing it in direct competition with Anthropic’s Claude and its recently launched Cowork application, which targets similar productivity use cases.
Context Window and Cost Structure
GPT-5.4 supports up to 1 million tokens of context in the API and Codex. There is a pricing threshold: input beyond 272,000 tokens is charged at double the standard cost per million tokens.
Tool Search
As AI tool ecosystems expand, loading every tool definition into a single prompt becomes expensive. GPT-5.4 addresses this with a new tool search feature in the API. The model receives a lightweight list of available tools and retrieves full definitions only when a specific tool is actually required. OpenAI frames this as a structural efficiency fix, reducing cost, latency, and context overhead simultaneously.
The pace of releases signals that OpenAI is prioritizing agentic, multi-step work over single-turn chat responses. Two major model updates in two days, each building on the last, points to an accelerated deployment cadence that shows no sign of leveling off.
Photo by BoliviaInteligente on Unsplash
This article is a curated summary based on third-party sources. Source: Read the original article
