GPT-5.4: Native Computer Use Now Beats Human Benchmarks

OpenAI has released GPT-5.4, OpenAI’s most capable general-purpose frontier model, arrived just two days after shipping GPT-5.3 Instant. The pace alone signals something. But GPT-5.4 is not an incremental refresh. It arrives with native computer use capabilities that, on at least one key benchmark, now exceed human performance on desktop navigation. Combined with a steerable reasoning mode, a 47% token efficiency gain on tool-heavy workflows, and deep integrations into Microsoft Excel and Google Sheets, GPT-5.4 represents the clearest indication yet that OpenAI is building toward a world where AI agents operate autonomously across your entire software stack.

The release comes in three configurations: standard GPT-5.4 (rolling out across ChatGPT, Codex, and the API), GPT-5.4 Thinking (available to Plus, Team, and Pro subscribers), and GPT-5.4 Pro (reserved for Enterprise and Edu users, as well as API access). GPT-5.4 Thinking will replace GPT-5.2 Thinking, which is now on a retirement clock ending June 3, 2026.

Native Computer Use: The Benchmark That Matters

The single most consequential capability in GPT-5.4 is native computer use. Unlike earlier approaches that bolted on browser automation as an afterthought, GPT-5.4 was built with computer use as a core capability. It can write code using libraries like Playwright to programmatically operate a computer, and it can also issue raw mouse and keyboard commands in response to screenshots. These two fundamentally different interaction modes together give it broad compatibility across desktop and web environments.

The benchmark numbers are hard to dismiss. On OSWorld-Verified (a test of desktop navigation using screenshots plus keyboard and mouse actions), GPT-5.4 scored 75.0% success. GPT-5.2, its predecessor, scored 47.3%. Reported human performance on the same benchmark sits at 72.4%. The model crossed the human threshold.

On BrowseComp, which measures how well an AI agent can persist across multi-step web research to locate hard-to-find information, GPT-5.4 Pro reached 89.3%, described by OpenAI as a new state of the art and a 17 percentage point absolute improvement over GPT-5.2. On Online-Mind2Web, the model achieves 92.8% success using screenshot-based observations alone.

These are not cherry-picked narrow tests. They measure the kind of multi-step, multi-application browsing and navigation that has historically separated AI tools from AI agents. GPT-5.4 closes most of that gap.

The Thinking Model Gets Steerable Reasoning

GPT-5.4 Thinking introduces a behavioral change that matters more than it initially appears: before responding to complex queries, it presents an outline of its reasoning plan. Users can interrupt, redirect, or refine that plan before the model commits to executing it.

“This makes it easier to guide the model toward the exact outcome you want without starting over or requiring multiple additional turns,” OpenAI says. That sounds like a minor UX improvement. It is actually a different model of human-AI collaboration: the user operates as a supervisor reviewing a plan rather than a prompter trying to steer an opaque process after the fact.

The practical consequence is that GPT-5.4 Thinking is better suited to tasks where first-pass accuracy matters: legal document drafting, financial modeling, code architecture design. In each of these domains, the cost of a wrong output is high enough that mid-process correction is worth the friction. The feature is live in the ChatGPT web app and on Android, with iOS support pending in a subsequent update.

On factuality, OpenAI reports that individual claims from GPT-5.4 are 33% less likely to be false compared to GPT-5.2. Given that hallucination rates have been one of the primary blockers for enterprise adoption of AI in high-stakes workflows, that improvement is worth tracking in real-world deployment.

Tool Search and the 47% Efficiency Claim

As AI agents grow more capable, they also grow more expensive to run. The problem is structural: modern agentic pipelines expose dozens or hundreds of tool definitions to the model on every request, even when only a handful are relevant. That context pollution drives up token costs and latency on every call.

GPT-5.4 addresses this with tool search, a new API feature that changes the model’s relationship to its tool ecosystem. Instead of receiving all tool definitions upfront, the model sees a lightweight list and retrieves full tool definitions only when it actually needs them. The result is a model that reasons about which tools are relevant before paying the cost of loading them.

OpenAI tested this on 250 tasks from Scale’s MCP Atlas benchmark with 36 MCP servers enabled. In the tool-search configuration, total token usage dropped 47% compared to the configuration that exposed all MCP functions directly in context, with no accuracy penalty. That 47% figure is specific to this evaluation design, not a blanket efficiency claim across all task types. But for anyone running production agentic systems at scale, a near-halving of token cost on tool-heavy workflows is a meaningful operating expense reduction.

This also positions GPT-5.4 well for the expanding ecosystem of autonomous AI agents, where models are expected to orchestrate dozens of specialized tools across long-running tasks.

Spreadsheets, Documents, and Enterprise Integration

GPT-5.4 ships with direct integrations into Microsoft Excel and Google Sheets, allowing the model to be plugged into individual cells and spreadsheet workflows for granular analysis and automated task completion. This follows similar moves by Anthropic, which has been building Claude integrations for financial spreadsheet workflows.

The practical use case is familiar to anyone who has spent time in finance or operations: instead of exporting data, feeding it to a model, and re-importing results, GPT-5.4 becomes a native participant in the document itself. It can interpret, transform, and generate structured data without leaving the spreadsheet environment.

The document and presentation capabilities extend the same logic to word processors and slide decks. OpenAI describes GPT-5.4 as capable of “generating spreadsheets, documents and presentations, and requiring less back-and-forth with a user.” The reduced back-and-forth is a meaningful capability claim: the model can parse ambiguous instructions and make reasonable judgment calls about intent rather than requesting clarification at every step.

These integrations, layered on top of native computer use, put GPT-5.4 in direct competition with enterprise automation platforms and, more pointedly, with a range of white-collar workflows that have previously required human judgment to execute. The question is no longer whether AI can perform these tasks, but how quickly enterprises will restructure workflows around the assumption that it can.

The 1 Million Token Context Window and Its Catch

GPT-5.4 supports up to 1 million tokens of context in the API and Codex. At that scale, a single model call can hold an entire codebase, a year of email threads, or a full year of financial filings, enabling agents to plan, execute, and verify tasks across genuinely long time horizons without losing context.

There is a catch. OpenAI charges double the standard per-million-token rate for input that exceeds 272,000 tokens. The 1M context ceiling is technically real; the economics of regularly using it are not always practical. For most workloads, the extended context is a capability that activates for high-value, long-horizon tasks rather than routine operations.

This mirrors the dynamic OpenAI has built across its developer toolchain, where flagship capabilities are technically available but priced to concentrate usage at the top of the market.

On document understanding, GPT-5.4 also extends its image input capabilities to an “original” detail level supporting up to 10.24 megapixels, and OmniDocBench performance improved to an average error of 0.109 from 0.140 for GPT-5.2, a meaningful gain for document extraction and OCR-adjacent use cases.

What GPT-5.4 Signals About the Agent Race

The release cadence alone is instructive. GPT-5.3 Instant shipped on March 3. GPT-5.4 shipped on March 5. OpenAI is not pacing releases around marketing calendars or conference schedules. It is shipping when capabilities are ready, at whatever speed the research pipeline allows.

The substance of GPT-5.4 points to where that pipeline is pointed: autonomous agents that operate software on behalf of users, reason transparently enough to be supervised, and do so at a cost efficiency that makes production deployment viable. The desktop navigation benchmark crossing the human performance threshold is not just a number: it means an AI agent running GPT-5.4 can now navigate a computer more reliably than a typical human user in controlled conditions.

That does not mean autonomous agents are ready to replace knowledge workers wholesale. The gap between benchmark performance and reliable, unsupervised real-world execution remains significant. Models still hallucinate, misinterpret context, and fail in edge cases that any experienced human would handle by instinct. But the trajectory is clear.

GPT-5.4 is not the end state. It is the current frontier of a race that all the major labs are running simultaneously, and OpenAI’s two-release week suggests it intends to move faster than the competition can comfortably track.