OpenAI Launches GPT-5.4: More Efficient, with Native Computer Use and a Professional Work Focus

Summary

On March 5, 2026, OpenAI launched GPT-5.4, which it describes as its "most capable and efficient frontier model for professional work." The model combines the reasoning advances of GPT-5.2 with the coding capabilities of GPT-5.3-Codex, while also expanding support for spreadsheets, documents, and presentations. It is available in three versions: GPT-5.4 (standard), GPT-5.4 Thinking (reasoning), and GPT-5.4 Pro (high performance).

The model is rolling out gradually in ChatGPT and Codex, and via the API. In ChatGPT, GPT-5.4 Thinking is available to Plus, Team, and Pro subscribers, replacing GPT-5.2 Thinking, which will be retired in three months. GPT-5.4 Pro is available via the API and for Pro and Enterprise plans. GPT-5.3 Instant remains the default model for regular conversations in ChatGPT.

In practice

OpenAI highlights six main areas of improvement with this release:

**Factuality:** GPT-5.4 is described as OpenAI's most accurate model to date — full responses are 18% less likely to contain errors, and individual claims are 33% less likely to be false compared to GPT-5.2.

**Token efficiency:** The model solves problems with significantly fewer tokens than its predecessors — in some cases up to 47% fewer — translating to faster responses and lower costs, even though the per-token price in the API is higher than GPT-5.2 to reflect improved capabilities.

**Context window:** GPT-5.4 supports up to 1 million tokens of context in the API — the largest context window OpenAI has ever offered — making it easier to analyse entire codebases, large document collections, or extended agent trajectories in a single call.

**Native computer use:** This is the first OpenAI "mainline" model with built-in computer-use capabilities, enabling agents to interact directly with software to complete, verify, and fix tasks in a build-run-verify-fix loop. In internal evaluations with HOA and property tax portals, GPT-5.4 achieved a 95% first-attempt success rate and 100% within three attempts, completing sessions roughly 3x faster and using approximately 70% fewer tokens than prior computer-use models.

**Tool Search (API):** A new feature that allows the model to receive a lightweight list of available tools and look up definitions only when needed, rather than including all tool definitions in the initial prompt. In systems with many tools, the previous approach could add tens of thousands of tokens to every request. The new approach dramatically reduces the tokens required and preserves the cache, making requests faster and cheaper.

**GPT-5.4 Thinking with upfront plan:** Similarly to Codex, GPT-5.4 Thinking now presents a summary of its working plan at the start of longer, more complex queries. Users can also provide instructions or adjust the model's direction mid-response, without needing to start over. This feature is available now on chatgpt.com and the Android app, with iOS coming soon.

On benchmarks, GPT-5.4 reaches 83% on GDPval — OpenAI's internal evaluation testing performance on knowledge work tasks across 44 occupations — up from 70.9% for GPT-5.2. The model also surpasses predecessors on the OSWorld-Verified and WebArena Verified computer use benchmarks, and beats the recent GPT-5.3-Codex release (and Google's Gemini 3.1 Pro) on the SWE-Bench Pro coding benchmark.

GPT-5.4 Pro, the higher-priced tier, stands out especially in advanced mathematics, scoring 38% on the FrontierMath benchmark for its hardest problems, compared to 27.1% for the Thinking version.

Context

OpenAI simultaneously published new safety research on monitoring model reasoning, including an open-source evaluation to test whether systems can conceal their chain-of-thought. Results indicate that GPT-5.4 Thinking shows a low ability to hide its reasoning, which the company characterises as a positive safety signal.

The release comes amid growing competitive pressure, with reports of users migrating from ChatGPT to rivals such as Anthropic's Claude. OpenAI maintains the same high cyber-risk classification used for GPT-5.3-Codex and has deployed additional protections, including expanded cyber safety systems, monitoring tools, trusted access controls, and request blocking for higher-risk activity.

Why it matters

• GPT-5.4 represents OpenAI's most direct push into the professional and enterprise work market, with measurable improvements in factuality, efficiency, and autonomous agency — areas where competition with Google and Anthropic is fiercest.
• The combination of native computer use, Tool Search, and a 1M token context window creates the technical conditions for more autonomous and reliable AI agents in real enterprise environments.
• The 33% reduction in the likelihood of false claims and 18% reduction in responses containing errors matters for professional use cases where accuracy is critical — though the evaluations are based on OpenAI's own internal benchmarks.
• The simultaneous publication of research on reasoning transparency signals that OpenAI is attempting to address structural safety concerns as models become more autonomous.