OpenAI has released GPT-5.4, its revolutionary frontier model that can be used on computers for the first time. With this version, AI agents may employ mouse, keyboard, and visual inputs to communicate directly with desktop apps, websites, and operating systems. This lets them automate complicated multi-step procedures without any help from people.
Core Features and Advancements
GPT-5. 4 supports a massive 1.05 million token context window, compared with the 400K of former versions like GPT-5. 3, which allows it to process large codebases or documents, or run long tasks fast. It also brings extreme reasoning modes (reasoning. (Work output of: effort: low, medium, high, xhigh) was optimized for research-grade reasoning, coding and professional work including but not limited to data analysis across spreads & presentations.
The model’s native computer use works in a “build-run-verify-fix” loop: it reads screenshots, issues commands and self-corrects errors, outperforming humans on benchmarks like OSWorld-Verified (75 percent compared to 72.4 percent human) and BrowseComp (82.7 percent). Other tools like web, file, code interpreter, tool and structured outputs, integrated for doing agentic stuff.
Pricing and Availability
Released for ChatGPT (Plus, Team, Pro), the API, first in Codex: GPT-5 4 costs $2.50 per million input tokens and $15 per million output tokens—more expensive than GPT-5. — 2’s input of $1.75, for instance, given that it has enhanced capabilities, but uses about 47% fewer tokens overall because it is designed to be more efficient Pro variants like GPT-5. 4 Pro offers even higher RPS limits, ranging from 500K TPM (Tier 1) to 40 MTPLM (Tier 5), tiered accordingly.
Benchmarks and Performance
| Benchmark | GPT-5.4 Score | Comparison |
| OSWorld-Verified (computer use) | 75% | Beats human 72.4% |
| BrowseComp (web reasoning) | 82.7% | State-of-the-art |
| Tool yields (efficiency) | 75% at 15 tools | Tops GPT-5.2’s 50% at 42 tools |
These improvements make GPT-5.4 perfect for running on its own for several hours, cutting down on mistakes in agent processes by 33%.
Implications for Developers and Businesses
For AI agents and developers: GPT-5. 4 keeps brittle wrappers from getting in the way of computer interaction, allowing desktop tasks like pulling spreadsheet info to fill out reports or control software to be seamlessly automated together. Enterprise and productivity are primed for a revolution in knowledge work, as we can already see AI making tools like Microsoft Office or enterprise software itself, transformed by background AI agents.
This launch amps up the competition with Claude (which has similar capabilities) and Google’s Project Mariner, driving an overall acceleration to agentic AI. Early adopters say it’s 2-3x faster at completing complex tasks, but safety safeguards are still vital to maintain control of the autonomy.