GPT-5.4 and Computer Use: OpenAI Enters the Autonomous Agents Era

OpenAI has just released GPT-5.4, which it describes as its most capable and efficient frontier model for professional work. With native computer use, a context window of up to 1 million tokens in the API and Codex, and improved agentic workflows, OpenAI fully enters the autonomous agents era. It ships in three variants: GPT-5.4, GPT-5.4 Thinking in ChatGPT, and GPT-5.4 Pro for the most demanding tasks.

Native Computer Use: AI That Controls Your Desktop

GPT-5.4 is OpenAI’s first general-purpose model with built-in computer use: it can operate software from screenshots, issuing mouse and keyboard commands, and also by writing code that drives applications with libraries like Playwright. In practice, an agent built on the model can:

Navigate web pages and complete forms
Operate desktop applications
Analyze interfaces from screenshots
Chain actions across multiple applications in a single workflow

On the OSWorld-Verified benchmark, which measures the ability to complete desktop tasks through screenshots and keyboard/mouse actions, GPT-5.4 reaches 75.0%, versus 47.3% for GPT-5.2 and above the human reference performance (72.4%).

┌──────────────────────────────────────────────────────┐
│   GPT-5.4 COMPUTER USE - FLOW (illustrative example) │
├──────────────────────────────────────────────────────┤
│                                                       │
│   User: "Extract CRM data and generate report"       │
│                                                       │
│   1. Opens browser → Accesses CRM                    │
│   2. Navigates to reports section                     │
│   3. Filters data by date and category                │
│   4. Exports CSV                                      │
│   5. Opens spreadsheet → Imports data                 │
│   6. Generates charts and formatting                  │
│   7. Saves final report                               │
│                                                       │
└──────────────────────────────────────────────────────┘

Unlike traditional API integrations, computer use allows AI to operate on existing software without custom connectors or development work, although supervising its actions is still recommended.

1 Million Tokens: More Context Than Ever

In the API, GPT-5.4 supports context windows of up to 1,000,000 tokens, the largest OpenAI has offered, compared with GPT-5.2’s 400K. This means:

Capacity	Approximate equivalent
1M tokens	~750,000 words
Source code	Large repositories in a single session
Documents	Many long documents at once
Conversation	Much longer work sessions

For development teams, this reduces the need to fragment context in long sessions. OpenAI also highlights better efficiency: GPT-5.4 solves the same problems with significantly fewer tokens than its predecessor.

Agentic Workflows: Codex and Automated Development

GPT-5.4 is available in Codex, OpenAI’s coding agent, taking AI-assisted development to the next level. A typical agentic flow with these tools can cover:

Assisted Development Pipeline

Requirements analysis - Interprets the task description
Planning - Proposes a plan and breaks it into steps
Implementation - Writes coordinated code across multiple files
Testing - Generates and runs tests
Code review - Reviews the changes looking for issues
PR preparation - Proposes changes with detailed descriptions

┌─────────────────────────────────────────────────────────┐
│           CODEX + GPT-5.4 AGENTIC PIPELINE              │
├─────────────────────────────────────────────────────────┤
│                                                          │
│   Task → Plan → Code → Test → Review → PR               │
│     ↑                                        │           │
│     └───── Developer feedback ───────────────┘           │
│                                                          │
└─────────────────────────────────────────────────────────┘

Improvements for Professional Work and Agents

Beyond computer use, the release brings relevant improvements for enterprise automation:

Fewer hallucinations: according to OpenAI, GPT-5.4 is 33% less likely to make errors in individual claims and 18% less likely to produce responses with errors than GPT-5.2
Tool Search in the API: the model looks up tool definitions only when needed, reducing tokens in systems with many tools
Knowledge work: 83% on GDPval, OpenAI’s evaluation of professional tasks
GPT-5.4 Thinking in ChatGPT shows a plan before starting complex tasks and lets you course-correct mid-response

Recall that Operator, the browser agent OpenAI launched in 2025, was folded into ChatGPT as ChatGPT agent in July of that year.

Comparison with Claude Computer Use

Anthropic was the pioneer: in October 2024, Claude 3.5 Sonnet became the first frontier model to offer computer use in public beta. The direct comparison:

Feature	GPT-5.4	Claude (Opus 4.6 / Sonnet 4.6)
Computer use	Native since GPT-5.4	Available since 2024 (beta)
Context	Up to 1M tokens (API and Codex)	1M tokens in beta (API)
Dev ecosystem	Codex	Claude Code + MCP
Chat availability	GPT-5.4 Thinking: Plus, Team, and Pro	Claude plans
Individual plan	ChatGPT Plus: $20/mo	Claude Pro: $20/mo

Claude maintains a very solid developer tooling ecosystem with Claude Code and MCP. GPT-5.4, for its part, arrives with a benchmark-leading score on OSWorld-Verified and broad integration across ChatGPT, the API, and Codex.

The reality is that both approaches are complementary. Many enterprise teams are adopting multi-model strategies, choosing the provider based on the use case.

Enterprise Implications and Use Cases

The most immediate applications for businesses include:

Process Automation

Accounting: Invoice processing, bank reconciliation
HR: Documentation management
Sales: CRM updates, lead follow-up
Support: Ticket resolution with access to internal systems

Software Development

Migrations: Updating legacy codebases with more context
Testing: Generation and execution of test suites
Documentation: Generation based on source code
DevOps: Infrastructure configuration and monitoring

Security and Control: The Elephant in the Room

Giving an AI access to your desktop and enterprise systems raises legitimate concerns. Before deploying agents with computer use, you should require:

Sandboxing: run agents in isolated environments with minimal permissions
Human approval: critical actions should require confirmation
Audit trail: a record of every action executed
Limits: restrictions by application and action type
Encryption: data protected in transit and at rest

The industry is still defining best practices, so the responsibility for designing these controls still falls on each organization.

How This Changes Software Development

GPT-5.4 with computer use transforms the relationship between developers and their tools:

The IDE as a secondary interface - Agents can operate directly on the editor
Automated visual testing - AI can verify UIs like a human QA would
End-to-end deploy - From commit to production, guided by AI and supervised by humans
Visual debugging - The agent can navigate the application, identify bugs, and propose fixes

This does not replace developers. Teams that adopt these tools will produce more with less friction, but human judgment in architecture, security, and UX remains fundamental.

Nextsoft Perspective

At Nextsoft, we see GPT-5.4 as a complementary tool, not a replacement for our current stack. Our strategy:

Claude Code remains our primary development tool for its precision and MCP ecosystem
GPT-5.4 with computer use is being evaluated for internal and client process automation
Computer use is ideal for visual testing and automated QA
1M context is useful for analyzing complete codebases in modernization projects

The key is using the right tool for each task, not betting everything on a single provider.

Conclusion

GPT-5.4 represents the moment where AI stops being a text generation tool and becomes an agent that operates in the digital world. With native computer use, massive context, and agentic workflows, OpenAI is redefining what enterprise automation means.

The competition between OpenAI and Anthropic benefits everyone. Each release pushes the boundaries of what is possible. For businesses and developers, the message is clear: autonomous agents are not the future, they are the present. The question is no longer whether to adopt them, but how to integrate them safely and effectively into your workflows.