OpenAI GPT-5.4 Beats Humans at Computer Use in 2026

OpenAI just shipped a model that operates a computer better than the median human tester. GPT-5.4 posted a 75.0% success rate on OSWorld-Verified, the benchmark that scores an agent's ability to drive a real desktop through screenshots and keyboard and mouse actions, edging past the 72.4% human baseline. Six months ago, GPT-5.2 managed only 47.3% on the same test. That single jump is the number that should make every software company recalculate its roadmap, because it changes what work a machine can be trusted to finish.

What Actually Happened

On launch day OpenAI introduced GPT-5.4 in three variants, standard, Pro, and Thinking, and called it its most capable and efficient frontier model for professional work. The headline capability is native computer use. In Codex and the API, GPT-5.4 is the first mainline general-purpose model that can operate a desktop environment directly, clicking, typing, scrolling, and navigating applications the way a person does, rather than calling a narrow set of predefined tools wired up by a developer in advance. The 75.0% OSWorld-Verified score is the public proof point, and it lands above the human reference for the first time in a shipping product rather than a research preview that never reaches customers.

The context window jumped to 1 million tokens, split as 922K input and 128K output. That is roughly the size of a long technical book held in working memory while the model plans, executes, and verifies a task across many steps without losing the thread. OpenAI paired the larger window with a feature it calls compaction, a mechanism that summarizes and prunes a long agent trajectory while preserving the context needed to keep going. GPT-5.4 is the first mainline OpenAI model trained to support compaction natively, which means it was designed from the start to run long sessions rather than to answer one question and forget everything.

The progression from 47.3% to 75.0% on OSWorld in roughly two release cycles is the steepest climb the benchmark has recorded since it was introduced. OpenAI is positioning the model squarely at agentic work: filling forms, operating internal tools, moving data between applications, and completing the kind of multi-window office tasks that have resisted automation because they require sustained, stateful attention rather than a single clever answer. The company framed the release around what the model can complete, not how it scores on reasoning exams, a deliberate shift in how a frontier launch is sold to the people who sign the checks.

Why This Matters More Than People Think

For two years the agent story was mostly demos. Models could write code and draft emails, but handing one a live desktop and walking away produced flaky, half-finished results that needed a human babysitter. A model that clears the human OSWorld baseline changes the unit of work a company can delegate. The task is no longer "answer my question" but "go do this in my software and come back when it is done." That shifts AI from an assistant that drafts to a worker that completes, and completion, not eloquence, is what enterprises actually pay for. The moment a model can be trusted to finish, the budget moves from software line items to labor line items.

The 1 million token window matters because agentic failures are usually memory failures. An agent that forgets step three by step nine loops, repeats actions, or corrupts the state of the document it is editing. Compaction plus a million tokens means the model can hold a full project brief, the relevant files, and its own running log of what it has already tried, all at once. That is the difference between an agent that finishes a forty-step workflow and one that quietly gives up at step twelve, and forty-step workflows, the reconciliations and migrations and multi-system updates, are exactly where the labor savings live and where prior models fell apart.

There is a pricing consequence too. If one GPT-5.4 session can complete a task that previously took a human knowledge worker an hour, the relevant comparison is no longer GPT-5.4 versus GPT-5.2. It is GPT-5.4 versus a salaried employee doing that hour of work. OpenAI is clearly aware of this framing, which is why it described the model in terms of professional work rather than raw intelligence. The benchmark that matters to a chief financial officer is cost per completed task, and native computer use is the capability that finally makes that number computable across the long tail of software no vendor ever bothered to build an API for.

The deeper shift is in who controls the integration layer. For a decade, automating a business process meant paying a systems integrator to wire APIs together, a project measured in months and six-figure budgets that locked companies into brittle pipelines. A model that operates the interface directly collapses that cost toward zero, because the integration is now just the agent looking at the screen and acting. That threatens an entire services industry built on connecting software that was never designed to talk, and it explains why OpenAI led with computer use rather than another reasoning record: the prize is not a smarter answer, it is the multi-billion-dollar market for making existing software finally work together without a custom build.

The Competitive Landscape

Anthropic introduced computer use first, in late 2024, and has iterated on it through successive Claude releases, making this the one frontier capability where OpenAI was visibly behind for over a year. Google has pushed Gemini into Android and Chrome as an agent that can act across apps, and Microsoft has rebuilt Copilot around an agent mode that takes actions inside Office rather than just suggesting text. The race to operate the computer, not merely describe it, is now the central front of the model wars, and GPT-5.4 is OpenAI's bid to retake the lead on the one metric that defines whether an agent is a toy or a tool.

The historical parallel is the shift from command-line tools to graphical interfaces in the 1980s. The graphical interface did not make computers smarter, it made them operable by people who did not want to memorize syntax, and that unlocked a far larger market almost overnight. Native computer use is the inverse move: it lets a model operate the graphical interfaces that humans built for humans, without anyone writing a custom integration for each application. The total addressable surface is every piece of software with a screen, which is most of them, including the decades of legacy enterprise tools that will never get a modern API.

The competitive risk for OpenAI is that computer use is hard to keep proprietary for long. The benchmark is public, the technique is broadly understood inside every frontier lab, and Anthropic and Google have deep agent teams shipping fast. Whatever lead the 75.0% score represents is measured in months, not years. The durable advantage will come from distribution, Codex, the API, and ChatGPT's hundreds of millions of users, rather than from the raw capability, which competitors will match on their next release. OpenAI knows the moat is the install base and the workflow lock-in, not the benchmark, because the benchmark is precisely the thing rivals can copy.

Hidden Insight: The Benchmark Is the Business Model

The quiet story in GPT-5.4 is that OpenAI has stopped selling intelligence and started selling completed work. Every previous flagship was marketed on reasoning scores, exam results, or coding benchmarks that measured how smart the model sounded in isolation. OSWorld measures something different and far more commercial: did the agent actually finish the task inside the software. By leading its launch with a computer-use number rather than a math or science score, OpenAI is signaling that the product is no longer a smarter chatbot but a digital worker that will be billed against the labor it replaces, which is a different company than the one that launched ChatGPT.

This reframes the entire revenue model. OpenAI reportedly crossed $25 billion in annualized revenue, and the path to the next tranche is not more chat subscriptions but agent deployments priced against headcount. A company that pays one knowledge worker $80,000 a year will happily pay a fraction of that for an agent that completes the same desktop tasks at 75% reliability, then route the 25% of failures to a human reviewer. That hybrid model, where the agent does the bulk and a person handles exceptions, is the actual product GPT-5.4 enables, and it is far more lucrative than selling seats to people who type questions into a box.

The compaction feature is the tell that OpenAI is optimizing for long-horizon work over flashy single answers. Compaction only matters if you expect sessions to run long enough that context overflows, which means OpenAI expects GPT-5.4 to be left running on multi-step jobs, not pinged for one-off questions. Training a mainline model to prune and summarize its own trajectory is an infrastructure investment in agents that work for minutes or hours unattended, not for seconds. The roadmap encoded in that one feature is autonomy, and compaction is the unglamorous plumbing that has to exist before any of the autonomy demos become reliable enough to charge for.

However, the bear case is straightforward and worth stating plainly. A 75% success rate means one in four tasks fails, and in a desktop environment a failure is not a wrong sentence, it is a wrong click that can delete a file, send an email to the wrong recipient, or submit a form that triggers a payment. Critics argue that the gap between 75% and the reliability enterprises need for unsupervised action, somewhere north of 99%, is not a small last mile but a different problem entirely. Beating the human baseline on an average makes a great headline, yet the variance is where the lawsuits live, and no benchmark score tells you what the worst 1% of an agent's actions look like when nobody is watching the screen.

What to Watch Next

In the next 30 days, watch for OpenAI to publish or leak pricing for GPT-5.4 computer-use sessions, because the per-task cost determines whether the labor-replacement math actually closes for a buyer. Also watch the Codex adoption numbers: if developers route real workflows through native computer use rather than scripted tool calls, that is the earliest signal the capability is production-ready rather than a benchmark trophy. Track which enterprise design partners OpenAI names publicly, since the first regulated-industry deployments, in finance, healthcare, and insurance, will reveal how much human oversight the 75% reliability really requires before a compliance team will sign off on letting an agent touch live systems.

Over 90 days, the question is whether Anthropic and Google ship computer-use numbers that match or beat 75.0%, which would compress OpenAI's lead to nothing and turn the capability into table stakes for every frontier model. Watch for an OSWorld-Verified leaderboard reshuffle and for the first independent audits that test the model on tasks well outside the benchmark distribution. If GPT-5.4 holds 75% on novel real-world software it has never seen, the capability is durable and the labor math holds. If it collapses to 50% on unfamiliar apps with cluttered interfaces, the headline number was overfit to the benchmark and

The other 90-day signal is whether a serious incident makes the news: an agent that deleted production data, filed the wrong regulatory document, or moved money it should not have. The first such headline will do more to set enterprise adoption pace than any benchmark, because it converts the abstract 25% failure rate into a concrete liability that a board can see. How OpenAI and its customers handle that first public failure, with disclosure or with silence, will shape the trust curve for the entire category.

By the 180-day mark, the leading indicator is employment data in narrow desktop-bound roles: data entry, claims processing, back-office reconciliation, and document review. If GPT-5.4-class agents are real, hiring in those categories slows first, quietly, before any model announcement makes the news. The other marker is OpenAI's own disclosures ahead of a possible late-2026 public listing, where it will have to quantify how much revenue comes from agents that complete work versus chat that answers questions. That split, when it finally appears in a prospectus, is the truest measure of whether native computer use became a real business or stayed an impressive benchmark that never paid rent.

OpenAI stopped selling a smarter chatbot the day GPT-5.4 cleared the human baseline at operating the computer itself.

---