OpenAI released GPT-5.4 on March 5, calling it “our most capable and efficient frontier model for professional work.” The headline feature: native computer-use capabilities that let AI agents click, type, and navigate software autonomously.
It’s not just incremental. On the OSWorld-Verified benchmark for desktop automation, GPT-5.4 scores 75% - surpassing both its predecessor (47.3%) and human testers (72.4%). OpenAI is betting that computer use, not conversation, is where AI creates real value.
What Computer Use Actually Means
GPT-5.4 can observe screenshots, issue mouse and keyboard commands, and execute multi-step workflows across applications. Unlike previous models that required developers to build automation infrastructure, GPT-5.4 handles this natively.
The practical applications are immediate: navigating enterprise software, filling out forms, extracting data from applications that lack APIs, coordinating actions across multiple programs. Tasks that previously required custom Selenium scripts or RPA tooling can now be described in natural language.
In VentureBeat’s coverage, OpenAI positioned this as the “first general-purpose model with native, state-of-the-art computer-use capabilities.” That’s partially marketing - Anthropic’s Claude has offered computer use since October 2024 - but GPT-5.4’s benchmark scores suggest meaningful improvements in reliability.
Beyond raw performance, the model introduces “tool search”: rather than requiring developers to upload detailed specifications for every available tool, GPT-5.4 can discover and invoke tools dynamically. According to SiliconAngle, this reduces token consumption by 47% on some tasks while maintaining accuracy - a significant cost savings for enterprises running agents at scale.
The Variants: Thinking vs Pro
GPT-5.4 ships in three configurations:
GPT-5.4 Thinking is the default for ChatGPT Plus subscribers ($20/month). It includes the model’s full reasoning capabilities but with standard rate limits.
GPT-5.4 Pro targets complex, high-stakes tasks. Reserved for ChatGPT Pro ($200/month) and Enterprise customers, it offers maximum performance on extended reasoning and computer-use workflows.
GPT-5.4 Standard runs in the API at $2.50 per million input tokens and $15 per million output tokens. For context windows exceeding 272,000 tokens, pricing doubles to $5 input and $22.50 output.
The API supports up to 1 million tokens of context - OpenAI’s largest ever - enabling agents to plan, execute, and verify across extended sessions.
ChatGPT for Excel: The Enterprise Play
Alongside GPT-5.4, OpenAI launched ChatGPT for Excel in beta. The add-in embeds ChatGPT directly into spreadsheets, letting users create, edit, and analyze workbooks through natural language.
According to The Decoder, GPT-5.4 scored 0.873 on internal investment banking benchmarks - significantly ahead of Claude Opus 4.6 (0.641) on the same tasks. Financial modeling, scenario analysis, and data extraction appear to be specific optimization targets.
OpenAI is also rolling out integrations with financial data providers: FactSet, Moody’s, S&P Global, and LSEG. The pitch is clear - bring market data, company financials, and AI reasoning into a single workflow.
The Excel integration is currently available in the US, Canada, and Australia for Business, Enterprise, Pro, and Plus tiers. A Google Sheets version is planned.
How It Compares
The frontier model landscape as of March 2026:
| Capability | GPT-5.4 | Claude Opus 4.6 | Gemini 3.1 Pro |
|---|---|---|---|
| Computer Use | 75% OSWorld | Available | Available |
| Coding (SWE-bench) | Strong | Leader | Strong |
| Abstract Reasoning | 94.3% GPQA | Comparable | 94.3% GPQA |
| Context Window | 1M tokens | 200K tokens | 2M tokens |
| Input Pricing | $2.50/M | $15/M | $2/M |
No single model dominates. GPT-5.4 leads on computer use and professional knowledge work. Claude Opus 4.6 remains the coding benchmark leader with deeper adaptive reasoning. Gemini 3.1 Pro offers equivalent abstract reasoning at roughly 15x lower cost, with the largest context window at 2 million tokens.
The Security Question
Computer use amplifies both capability and risk. An AI that can click buttons can also click the wrong buttons.
GPT-5.4’s system card reveals that OpenAI classified the model as “high cyber-risk” - the same rating as GPT-5.3-Codex - and deployed additional safeguards. According to security analysts, GPT-5.4 Thinking is OpenAI’s first general-purpose model with “implemented mitigations for high cybersecurity capability,” rather than just flagging risks.
These mitigations include expanded monitoring tools, trusted access controls, and request blocking for “higher-risk activity” on zero-data-retention surfaces. The specifics remain vague - OpenAI hasn’t published detailed technical documentation on what triggers blocks or how access controls work.
The broader enterprise landscape is troubling. A recent Help Net Security survey found that 80% of organizations reported risky agent behaviors, including unauthorized system access and improper data exposure. Only 21% of executives claimed complete visibility into agent permissions, tool usage, or data access patterns.
Computer-use agents are proliferating faster than security practices can adapt. GPT-5.4 may be safer than its predecessors, but “safer” in an immature field doesn’t mean safe.
What This Means
GPT-5.4 represents a real capability jump. The benchmarks on computer use are impressive, the tool-search efficiency gains are meaningful, and the Excel integration targets legitimate enterprise workflows.
But the framing matters. OpenAI isn’t just releasing a smarter chatbot - it’s building infrastructure for autonomous software agents. The model exists in an API, which means developers are already wiring it into production systems where it can take actions with real consequences.
The questions that follow aren’t just technical:
Who’s liable when an AI agent clicks the wrong button in your accounting software?
How do you audit decisions made by a model that can observe screens, reason about options, and execute without human confirmation?
What happens when these capabilities become cheap enough that every SaaS product has an AI agent controlling other AI agents?
GPT-5.4 is a capable tool. Whether the systems we’re building around it are ready for that capability is a different question entirely.
The Bottom Line
OpenAI’s GPT-5.4 delivers on computer use, outperforming humans on desktop automation benchmarks while using fewer tokens per task. The Excel integration and financial data partnerships signal a serious enterprise push. But computer-use AI is moving faster than the security and governance practices needed to deploy it safely. The model works. The question is whether we’re ready for what it enables.