GPT-5.4: OpenAI Ships Native Computer Use and 1M Context

GPT-5.4, OpenAI’s latest frontier model released March 5, 2026, is the company’s most capable and efficient model to date. The GPT-5.4 release is not a minor iteration. It arrives as the first general-purpose model in OpenAI’s lineup with native computer-use capabilities, a 1-million-token context window, and benchmark results that mark a meaningful step toward autonomous professional agents. Two days after quietly launching GPT-5.3 Instant, OpenAI is signaling that the cadence of major model drops is accelerating.

What GPT-5.4 Actually Does Differently

The headline capability is computer use. In the API and Codex, GPT-5.4 can operate a computer by issuing keyboard and mouse commands in response to screenshots, navigate web browsers, and coordinate workflows across applications without requiring custom integration for each tool. This is not a demonstration feature; OpenAI is deploying it as the default recommended model for agentic workloads in Codex.

This matters because prior computer-use implementations, including earlier operator tools and third-party wrappers, required models to be explicitly scaffolded around screen capture and action loops. GPT-5.4 handles that natively, reducing the engineering overhead for building agents that interact with existing desktop and web software.

The 1-million-token context window is the second structural shift. Agents running long-horizon tasks, code review across entire repositories, document synthesis across hundreds of pages, or multi-step workflow planning now have far more headroom before hitting context limits. OpenAI does charge double per million tokens once input exceeds 272,000 tokens, which will matter at scale, but the ceiling is no longer the binding constraint it was.

Benchmark Numbers Worth Scrutinizing

OpenAI published several benchmark comparisons against GPT-5.2, its prior generation reasoning model.

On GDPval, which tests AI performance across 44 occupations spanning the top industries contributing to U.S. GDP, GPT-5.4 matches or exceeds industry professionals in 83.0% of comparisons. GPT-5.2 hit 70.9%. The tasks in GDPval are substantive: sales presentations, accounting spreadsheets, manufacturing diagrams, urgent care schedules. This is not multiple-choice trivia.

On internal spreadsheet modeling benchmarks calibrated to junior investment banking analyst work, GPT-5.4 scores 87.3% versus 68.4% for GPT-5.2. The BigLaw Bench legal analysis evaluation gives GPT-5.4 a 91% score. Mercor’s APEX-Agents benchmark, which tests performance on professional services work, places GPT-5.4 at the top of its leaderboard.

On hallucination reduction, OpenAI reports that individual claims from GPT-5.4 are 33% less likely to be false than GPT-5.2, and full responses are 18% less likely to contain any errors. Token efficiency has also improved significantly: on some tasks, GPT-5.4 uses 47% fewer tokens than predecessors, which translates to lower API costs and faster inference.

These are vendor-reported numbers, and independent replication will matter. But the directional shift they describe, toward models that are simultaneously more capable, more factual, and cheaper to run, is consistent with what third-party evaluators have observed across the GPT-5 series.

The Agentic Architecture Behind the Release

GPT-5.4 includes a feature called tool search, which helps agents identify and invoke the correct tool from a large ecosystem of connectors without degrading reasoning quality. For developers building on the OpenAI API with dozens or hundreds of available tools, this reduces a significant source of agent failure: models calling the wrong tool or hallucinating tool parameters.

The GPT-5.4 Thinking variant, rolling out to ChatGPT Plus, Team, and Pro subscribers, adds mid-response steering. Users can see the model’s plan before it finishes and adjust course without restarting the conversation. OpenAI describes this as reducing the number of turns required to reach the desired output, a practical improvement for complex iterative tasks.

OpenAI also launched a ChatGPT for Excel add-in alongside GPT-5.4, with direct integration into Google Sheets. These integrations allow GPT-5.4 to operate inside spreadsheet cells, run analysis, and automate tasks without requiring users to copy data into ChatGPT. Anthropic’s Claude Cowork, launched earlier this year, targets the same workflow. The competition for enterprise spreadsheet automation is now direct and active. As we covered in OpenAI’s push to own the full developer stack, the company is systematically closing the gaps between AI capability and the tools professionals already use daily.

What This Signals for the Agent Ecosystem

GPT-5.4 arriving with native computer use is not an isolated product decision. It reflects a broader architectural bet that agents will not primarily operate through custom APIs but through the same interfaces humans use: browsers, applications, command lines, and file systems. This makes the agent surface area vastly larger, and the value of general-purpose computer-use capability proportionally higher.

The inference economy is reshaping how enterprises think about AI deployment. With token efficiency now a key competitive axis, a model that delivers superior benchmark performance while consuming fewer tokens per task threatens to compress margins for competing providers simultaneously on quality and cost.

For developers currently building on Anthropic’s Claude or Google Gemini, GPT-5.4 raises the evaluation floor again. The three-month deprecation window for GPT-5.2 Thinking, which retires June 3, 2026, also forces a practical migration decision. OpenAI is not leaving prior-generation models in the catalog indefinitely.

The Remaining Questions

Native computer use at the model level does not solve the hard problems: reliability across long task sequences, graceful handling of unexpected UI changes, and audit trails for enterprise compliance. GPT-5.4 can issue keyboard commands, but whether it can sustain a 50-step workflow without failure in a production environment remains to be tested outside controlled benchmarks.

The pricing structure also deserves attention. The 272,000-token threshold at which costs double will affect long-context agentic applications more than short-turn conversational ones. Enterprise buyers designing agent pipelines around 1M context should model their actual usage distributions carefully before committing.

What is clear is that GPT-5.4 moves the baseline for what a frontier model is expected to do. Computer use is now table stakes for the leading general-purpose model. The companies that do not have a credible answer to it in the next cycle will feel that absence in enterprise evaluations.