GPT-5.4 Unveiled: 1M‑Token Context Window and Native Computer Control

Release Overview

On March 6, 2026, OpenAI announced GPT-5.4, positioning it as the most advanced model for professional work scenarios. The announcement emphasizes three simultaneous upgrades: a 1 million‑token context window, native computer‑control capability, and a substantial boost in factual accuracy.

Three Model Tiers

Standard – targets API developers and enterprise integration, offering a balanced performance and the full 1 million‑token context.

Thinking – includes a built‑in reasoning chain for deep analysis, aimed at Plus and Team subscribers.

Pro – delivers the highest performance ceiling, ranking first in legal and financial evaluations, for Pro and Enterprise users.

1 Million‑Token Context Window

The API now supports up to 1 million tokens, the largest context size OpenAI has provided. This capacity is roughly equivalent to a 700‑page novel, five years of a mid‑size company's financial statements, hundreds of legal contracts, or the core logic of a large code repository. Previously, long‑document tasks suffered from the "forgetting earlier context" problem, requiring manual chunking and re‑prompting; the new window largely eliminates that issue, delivering a tangible efficiency gain for lawyers, financial analysts, compliance officers, and researchers.

Native Computer‑Use Capability

GPT‑5.4 is the first OpenAI flagship model to integrate Computer Use natively. It can:

Open applications and navigate UI elements.

Perceive the current screen via screenshots and issue mouse‑click and keyboard commands.

Directly manipulate data in Microsoft Excel and Google Sheets (financial plugins pre‑installed).

Execute multi‑step, cross‑application workflows while maintaining full contextual continuity.

Benchmark results illustrate the impact:

OSWorld‑Verified (desktop screenshots + mouse/keyboard): 75.0% success, up from 47.3% in GPT‑5.2 (+27.7 points).

Online‑Mind2Web (pure screenshot web browsing): 92.8%.

BrowseComp (persistent web search) – Pro version scores 89.3%, a new state‑of‑the‑art record (+17% over the previous model).

Accuracy Improvements

OpenAI provides a side‑by‑side comparison with GPT‑5.2:

Single‑statement error probability reduced by 33%.

Overall response error proportion reduced by 18%.

GDPval professional knowledge test: 83 points.

APEX‑Agents legal and financial assessment: ranked first in the industry.

Token consumption per problem markedly lower, translating to lower API costs and faster responses.

Developer Feature: Tool Search

GPT‑5.4 adds a "Tool Search" mechanism. When building an AI Agent, the model no longer needs to embed every tool definition in the system prompt; it queries tool specifications on demand. Benefits include cleaner system prompts, reduced token usage, support for larger tool libraries, lower engineering complexity, and new possibilities for enterprise‑grade AI automation.

Competitive Landscape

Anthropic’s Claude series and Google’s Gemini are also pursuing computer‑use and enterprise integration. By embedding native Computer Use in its flagship model, OpenAI signals an intent to settle the competition decisively. A quoted strategic message reads, "In professional work scenarios we remain number one and will widen the gap." The three new capabilities—massive context, native computer control, and Office‑plugin integration—are aimed at enterprise procurement decision‑makers rather than hobbyist users.

Implications

The release marks a shift from AI as an experimental tool to a core component of digital workforces. GPT‑5.4’s performance sets a new benchmark, prompting Anthropic, Google, and domestic AI firms to accelerate their own developments. The race is no longer about who releases first, but who can become an indispensable productivity tool.