OpenAI GPT-5.4: Native Computer Use, 1M Token Context

OpenAI GPT-5.4 is the company’s most capable frontier model to date, released March 5, and the first general-purpose AI OpenAI has shipped with native computer-use built in. The model can click, type, and navigate software using screenshots and mouse-and-keyboard commands without relying on a separate, specialized agent layer. For the enterprise market, that distinction matters: it removes an integration step that has kept computer-use AI largely in the hands of developers willing to stitch together custom toolchains.

GPT-5.4 ships in two tiers. GPT-5.4 Thinking is available to all paid ChatGPT subscribers, starting at the $20-per-month Plus plan. GPT-5.4 Pro is reserved for ChatGPT Pro ($200 monthly) and Enterprise users, targeting the most computationally demanding workloads. Free ChatGPT users will receive access through auto-routing only. GPT-5.4 Thinking also replaces GPT-5.2 Thinking in the ChatGPT model picker immediately, with GPT-5.2 Thinking moving to a Legacy Models section and retiring on June 5, 2026. Both tiers are available in OpenAI’s API and its Codex software development platform.

Native Computer Use Changes the Agentic Equation

The headline capability is computer use embedded directly in the model rather than bolted on as an afterthought. GPT-5.4 can both write code to operate software through libraries such as Playwright and issue raw mouse and keyboard commands in response to screenshots, operating across applications the way a human operator would. OpenAI is making the feature available through the API and its Codex platform, the software development environment the company has been building as a developer-first interface for agentic workflows.

Benchmark results show meaningful and verifiable progress over prior versions. On BrowseComp, which measures an AI agent’s ability to persistently browse the web to locate hard-to-find information, GPT-5.4 improved by 17 percentage points absolute over GPT-5.2. GPT-5.4 Pro reached 89.3 percent on BrowseComp, a new state-of-the-art result at the time of release. The model also set records on OSWorld-Verified and WebArena Verified, two standard evaluations for desktop navigation and web-based task completion respectively.

OpenAI reports that GPT-5.4 produces 33 percent fewer false claims and 18 percent fewer errors overall compared to GPT-5.2. On its GDPval test, which measures performance across knowledge work tasks spanning 44 professional occupations, GPT-5.4 scored 83 percent. OpenAI has also published the chain-of-thought controllability evaluation methodology as open source, a move that invites external scrutiny rather than asking the research community to take benchmark numbers at face value.

A Million Tokens and a 47 Percent Efficiency Jump

GPT-5.4 supports context windows of up to 1 million tokens in the API and Codex, enabling agents to hold large codebases, extended conversation histories, or substantial datasets within a single interaction without chunking or summarization workarounds. OpenAI charges double the standard rate per million tokens once input exceeds 272,000 tokens, a pricing structure that will influence how developers architect long-context workflows and whether the full window is worth the additional inference cost.

Efficiency is the less visible but commercially significant improvement in this release. OpenAI reports that GPT-5.4 uses 47 percent fewer tokens than its predecessors on some task categories. That reduction lowers inference costs for API-scale deployments and makes the model faster for real-time applications where latency matters. At production scale, a nearly 50 percent token reduction translates directly into infrastructure savings, which is relevant for any organization running high-volume agentic pipelines.

The model ships with a new suite of ChatGPT integrations that plug directly into Microsoft Excel and Google Sheets, allowing GPT-5.4 to read and write individual spreadsheet cells and execute granular analysis tasks within those environments. The move follows similar integrations from Anthropic’s Claude and positions OpenAI to compete for finance and operations workflows at the spreadsheet level, where the practical value of AI assistance is immediate and measurable rather than abstract.

Competitive Context and What This Means for the Frontier

GPT-5.4 arrives during the most compressed period of frontier model competition the industry has seen. The release came just two days after GPT-5.3 Instant, itself a model OpenAI said reduced hallucinations by 26.8 percent. The cadence suggests OpenAI is shipping iterative improvements at a pace designed to prevent competitors from holding any meaningful benchmark lead for long.

The native computer-use capability in particular puts pressure on the broader agentic AI field to deliver equivalent functionality in general-purpose models rather than as a separate specialist system. The race for raw AI compute capacity continues to accelerate in parallel, with infrastructure investment at an all-time high across every major lab. That investment now needs models capable of acting autonomously in real environments, not just generating text.

The central open question is whether GPT-5.4’s benchmark performance translates to reliable autonomous operation in production environments. Computer-use AI has historically struggled with edge cases and error recovery when operating outside structured task definitions. OpenAI’s decision to release GPT-5.4’s computer-use mode broadly through the API, rather than in a gated research preview, is the clearest signal yet that the company believes the technology is ready to be tested at scale by developers who will find the failure modes quickly and publicly.