AI News

GPT-5.4: OpenAI moves to autonomous agents with computer use

OpenAI launches GPT-5.4 with native computer use, 1 million tokens of context and 33% fewer hallucinations. Analysis of the implications for businesses.

5 min read
OpenAIGPT-5.4Agents IAComputer UseLLM
âš¡ The news in 30 seconds

OpenAI launches GPT-5.4: the first general-purpose model able to use a computer

On March 5, 2026, OpenAI launched GPT-5.4, its most advanced model to date. The major new feature: native computer use — the model can navigate a browser, click buttons, fill in forms and run complete workflows across several applications. On the OSWorld-Verified benchmark (desktop navigation), GPT-5.4 scores 75%, surpassing human performance (72.4%). The model also features a 1.05 million token context window, reduces hallucinations by 33% compared to GPT-5.2 and introduces Tool Search, a feature that reduces token consumption on agentic tasks by 47%.

We're no longer talking about a chatbot that answers questions, but an agent able to act on a computer. For businesses, this is the shift from conversational AI to operational AI.

What it changes for you

✦ The opportunity

GPT-5.4 marks a turning point: for the first time, a general-purpose language model can perform tasks on a computer with a level of reliability higher than humans on some benchmarks. For SMBs and mid-market companies, this opens up three categories of immediately actionable use cases:

🖥️

Desktop workflow automation

Repetitive tasks on business software — entering data into an ERP, extracting information from a supplier portal, filling in administrative forms — can now be delegated to a GPT-5.4 agent. Unlike traditional RPA solutions that break with every interface update, computer use adapts visually to layout changes. It's intelligent RPA that doesn't require fragile scripts.

📊

83% performance on professional tasks

On the GDPval benchmark — which evaluates models across 44 professions — GPT-5.4 matches or exceeds human professionals in 83% of cases. In financial modeling (investment banking), the score rises from 68.4% (GPT-5.2) to 87.3%. For mid-market companies that hesitated to entrust complex analytical tasks to AI, this level of performance changes the cost/risk equation.

🔧

Tool Search: more efficient, lower-cost agents

The new Tool Search feature lets the model dynamically search for the right tools in a catalog, instead of loading all definitions into the prompt. The result: 47% fewer tokens on multi-tool tasks, so execution cost is almost halved for complex agents. For companies deploying agents with 10+ connected tools, this is a significant saving.

âš  The risk

⚠️

An agent that clicks can also make costly mistakes

Computer use gives the model the ability to act on a computer — including incorrectly. A click in the wrong place in an ERP, an erroneous order validation, a file deletion: the consequences can be immediate and irreversible. OpenAI classifies GPT-5.4 as “High cyber capability” in its Preparedness Framework. Companies must deploy computer use with strict safeguards: sandboxed environments, human validation before critical actions, exhaustive logging.

💰

Premium pricing for the Pro version

While GPT-5.4 standard remains affordable ($2.50/$15 per million tokens), the Pro version at $30/$180 per million tokens places maximum performance at a price 12× higher. For use cases requiring the best quality (complex legal analysis, financial modeling), the AI budget can quickly balloon. Assess whether the standard version is enough before switching to Pro — in 80% of cases, it is.

Our recommendation

GPT-5.4 is a significant advance, but adoption must be methodical. Here is our three-step action plan:

1

Identify your repetitive desktop workflows

List the tasks your teams perform daily on web or software interfaces: entering data into a CRM, extracting data from a portal, copy-pasting between applications, filling in administrative forms. Every task that follows a predictable pattern is a candidate for computer use. Prioritize those that consume the most time and carry the least risk in case of error.

2

Test GPT-5.4 vs your current stack

Run a comparative benchmark on your 3 priority use cases: GPT-5.4, Claude Opus 4.6 and your current solution (manual or scripted). Measure output quality, execution time, cost per task and error rate. The Thinking version of GPT-5.4 (available to Plus, Team and Pro subscribers) offers an advanced reasoning mode useful for complex analytical tasks.

3

Deploy with mandatory human oversight

For the first computer use deployments, require human validation before every write or modification action. Use a sandboxed environment (VM or container) to limit the scope of errors. Log every agent action for audit. Once reliability is validated over 500+ executions without a critical error, you can gradually reduce oversight on low-risk tasks.

In summary

Opportunity
Native computer use, 83% on professional tasks, 47% tokens saved with Tool Search
Risk
Execution errors on desktop, high Pro pricing, dependence on OpenAI
Recommended action
Identify desktop workflows, benchmark vs Claude/Gemini, deploy with oversight
Horizon
Immediate — GPT-5.4 available via API and ChatGPT right now

Frequently asked questions

What is computer use in GPT-5.4?

Computer use lets the model operate a computer directly: navigate a web browser, click interface elements, fill in forms, run multi-step workflows across several applications. GPT-5.4 is OpenAI's first general-purpose model to integrate this capability natively. On the OSWorld-Verified benchmark, it scores 75%, surpassing human performance (72.4%).

Is GPT-5.4 better than Claude Opus 4.6?

Both models excel in different areas. GPT-5.4 dominates on professional tasks (83% on GDPval) and computer use (75% on OSWorld). Claude Opus 4.6 retains an advantage on specialized coding (Terminal-Bench) and offers a 1-million-token context window with its extended thinking mode. The choice depends on the use case: desktop agents for GPT-5.4, document analysis and code for Claude.

What is the cost of GPT-5.4 compared to GPT-5.2?

GPT-5.4 standard costs $2.50 per million input tokens and $15 output, versus $1.75/$14 for GPT-5.2. The 43% input premium is offset by greater efficiency: the model consumes 47% fewer tokens on agentic tasks thanks to the Tool Search feature. Net, the cost per task is often lower. The Pro version ($30/$180) targets complex cases requiring maximum performance.

Can GPT-5.4 be used to automate business tasks right now?

Yes, via the API or Codex. The most mature use cases are automating web workflows (data entry, structured scraping, interface testing), generating reports from multiple sources and code assistance with visual debugging. For critical tasks, maintain human oversight — computer use is powerful but not infallible.

For technical profiles

The GPT-5.4 variants

GPT-5.4 Standard

The all-rounder model

Context window of 1.05 million tokens. Native computer use for desktop automation. 33% fewer hallucinations than GPT-5.2. Tool Search for dynamic tool management. 83% performance on GDPval (44 professions). Ideal for enterprise agents, document analysis and workflow automation.

GPT-5.4 Thinking

The advanced reasoning mode

Version with extended chain of thought for complex reasoning tasks: financial analysis, multi-step problem solving, strategic planning. Available to ChatGPT Plus, Team and Pro subscribers. Higher latency but superior reasoning quality on analytical tasks.

GPT-5.4 Pro

Maximum performance

Premium version for the most demanding use cases. Score of 87.3% in financial modeling (vs 68.4% for GPT-5.2). Pricing 12× higher than the standard model. Reserved for critical tasks requiring maximum reliability: audit, legal analysis, complex modeling.

GPT-5.4 API pricing

Standard input $2.50/M tokens
Standard output $15/M tokens
Pro input $30/M tokens
Pro output $180/M tokens

Comparison of frontier models (March 2026)

CriterionGPT-5.4Claude Opus 4.6Gemini 3.1 ProDeepSeek V4
Computer useNative (75% OSWorld)Via third-party APILimitedNo
Max context1.05M tokens1M tokens1M tokens256k tokens
Professional tasks (GDPval)83%Not measuredNot measuredNot measured
Code (SWE-bench)~78%72%80.6%Not disclosed
Input cost ($/M)$2.50$15$2$0.50
Open-weightNoNoNoYes

Related articles