OpenAI GPT-5.4: Native Computer Use and Excel Integration

OpenAI has released GPT-5.4, its most capable general-purpose model to date, introducing native computer use capabilities alongside a major token efficiency leap and new integrations that bring the model directly into enterprise spreadsheets. OpenAI’s official release notes confirm GPT-5.1 models were retired on March 11, 2026, with GPT-5.4 becoming the primary model across ChatGPT plans. The announcement follows a rapid succession of model updates, with GPT-5.3 Instant launching just two days prior, signaling that OpenAI is accelerating its release cadence as competition from Anthropic, Google DeepMind, and xAI intensifies.

The release simultaneously retired GPT-5.1 models across ChatGPT on March 11, 2026, and sets a June 5 expiration for GPT-5.2 Thinking, a sign that OpenAI is aggressively collapsing its legacy model surface area as the GPT-5 family matures.

Two Variants, Sharply Differentiated by Use Case

GPT-5.4 ships in two configurations. GPT-5.4 Thinking is available to all paid ChatGPT subscribers starting at the Plus tier ($20/month) and through the API. GPT-5.4 Pro, designed for the most computationally demanding tasks, is restricted to ChatGPT Pro ($200/month) and Enterprise plan holders. Both variants are also available within Codex, OpenAI’s software development environment.

Free tier users will receive limited access through auto-routing, where the model handles queries that fall within its operational scope without the user selecting it directly.

The efficiency numbers are the headline from an infrastructure standpoint. OpenAI claims GPT-5.4 uses 47% fewer tokens on certain task classes compared to its predecessors, a figure that has significant downstream implications for API cost at scale. The model also supports up to 1 million tokens of context through the API and Codex, though OpenAI charges double the standard per-token rate once inputs exceed 272,000 tokens.

Native Computer Use Closes the Gap with Autonomous Agents

The capability drawing the most industry attention is GPT-5.4’s native computer use mode, available through the API and Codex. This is OpenAI’s first general-purpose model released with what the company describes as state-of-the-art computer use built in, not as an add-on or separate product, but as a core model capability.

The model can write code to operate computers via libraries like Playwright and issue direct mouse and keyboard commands based on screenshot input, allowing agents to navigate desktop environments, switch between applications, and complete multi-step workflows without human intervention at each transition.

Benchmark results offer some grounding for these claims. On BrowseComp, which measures persistent web browsing to locate hard-to-find information, GPT-5.4 shows a 17 percentage point absolute improvement over GPT-5.2. GPT-5.4 Pro reaches 89.3% on the same benchmark, which OpenAI positions as a new state of the art. Performance on OSWorld-Verified, measuring desktop navigation via screenshots and input commands, also improved, though OpenAI has not released the full numerical breakdown.

For developers building on the agentic AI infrastructure that has been advancing rapidly through early 2026, the native computer use capability removes a significant architectural hurdle. Prior approaches typically required wrapping a model with a separate computer use layer, introducing latency and additional failure points. Having the capability baked into the base model simplifies deployment and makes fully autonomous agent pipelines more practical for production environments.

Excel Integration Brings GPT-5.4 Into the Enterprise Workflow Core

Beyond raw model capability, OpenAI is shipping a suite of ChatGPT integrations that embed GPT-5.4 directly into Microsoft Excel cells and formulas, with Google Sheets support announced for later this quarter. The integration enables granular spreadsheet analysis and automated task completion within tools that remain the operational backbone of most large enterprises.

The timing is deliberate. Anthropic has been making coordinated moves in the enterprise software space, including its own Excel integrations via Claude and the recent launch of Claude Cowork. GPT-5.4’s spreadsheet plugins represent OpenAI’s direct counter, targeting the same finance, operations, and data analysis workflows that Anthropic has been pursuing.

The white-collar displacement question these tools raise is not hypothetical. OpenAI’s March 2026 launches, from GPT-5.3’s hallucination reduction to GPT-5.4’s autonomous computer operation and spreadsheet embedding, form a sequential capability stack that collectively covers a significant portion of knowledge worker task classes.

For developers evaluating the model against GPT-5.3 Codex, third-party comparisons position GPT-5.4 as the clear upgrade across benchmarks with data-intensive workloads, with GPT-5.2 retirement giving developers a hard June 5 deadline to migrate. As OpenAI’s developer tooling strategy continues to expand through Codex and API integrations, GPT-5.4 represents the new baseline for production workloads.

What Comes Next

OpenAI has not indicated when GPT-5.5 or a successor variant will arrive, but the two-day gap between GPT-5.3 and GPT-5.4 suggests release intervals are compressing. The simultaneous retirement of GPT-5.1 and the announced expiration of GPT-5.2 indicate the company is no longer maintaining wide backward compatibility windows.

For enterprise adopters, the immediate questions are practical: migration timelines for GPT-5.2 API users, cost modeling against the 272,000-token billing inflection point, and whether native computer use unlocks enough workflow automation to justify the Pro tier pricing at scale.