OpenAI launches GPT-5.4 with native computer use, stronger reasoning, and upgraded coding performance

OpenAI has released GPT-5.4, a new frontier model the company positions for professional work across reasoning, coding, and long-horizon agent workflows. OpenAI says GPT-5.4 rolls out in ChatGPT as GPT-5.4 Thinking and also ships in the API and Codex, with a higher-compute GPT-5.4 Pro option for complex tasks.

OpenAI says GPT-5.4 adds native computer-use capabilities in the API and Codex, letting agents operate computers across applications and workflows. The company also highlights a 1M-token context limit for long-horizon planning and verification.

For ChatGPT users, OpenAI says GPT-5.4 Thinking can provide an upfront plan so people can adjust course mid-response without starting over. OpenAI also says GPT-5.4 improves deep web research for highly specific queries while maintaining context during longer reasoning.

What’s new in GPT-5.4

• Upfront plan in ChatGPT (GPT-5.4 Thinking) designed to improve steerability during long reasoning.
• Native computer-use in the API and Codex for multi-app workflows that involve screenshots and UI actions.
• Higher token efficiency versus GPT-5.2, which OpenAI says reduces token usage and speeds up reasoning-heavy work.
• A Pro variant that uses more compute and targets harder tasks, offered through OpenAI’s Responses API.

Benchmarks OpenAI highlighted for GPT-5.4

OpenAI published the following headline results comparing GPT-5.4 to GPT-5.3-Codex and GPT-5.2.

OpenAI says GDPval measures “well-specified knowledge work” across 44 occupations, and it reports GPT-5.4 matches or exceeds industry professionals in 83.0% of comparisons on that evaluation.

Availability and rollout details

OpenAI says GPT-5.4 Thinking becomes available to ChatGPT Plus, Team, and Pro users starting March 5, 2026, and it replaces GPT-5.2 Thinking in the model picker. OpenAI also says GPT-5.2 Thinking remains available for three months under Legacy Models before retirement on June 5, 2026.

For developers, OpenAI documents GPT-5.4 Pro as a Responses API model that supports computer use, web search, file search, and tool search, along with a 1,050,000-token context window and a 128,000-token max output.

Practical impact for teams using GPT-5.4

If you build internal tooling, GPT-5.4’s combination of long context and native computer use can reduce the number of handoffs between separate models and scripts for tasks like QA flows, data checks inside spreadsheets, or multi-step browser work. OpenAI also calls out improved spreadsheet and presentation creation, including guidance and examples it links from its announcement.

If you manage risk, OpenAI’s GPT-5.4 Thinking system card describes the model as the first general-purpose model in the GPT-5 series with mitigations for “High capability in Cybersecurity.” That note matters for organizations that gate model access or apply stronger controls to cybersecurity-relevant use cases.

What to check before you switch production workloads

• Pin a snapshot ID for consistent behavior in production deployments, since OpenAI lists dated snapshot aliases for GPT-5.4 Pro.
• Validate pricing for your workload, especially if you plan to use long-context sessions or computer-use tool calls.
• Decide your default reasoning.effort level for Pro usage, since OpenAI documents medium, high, and xhigh options.
• Review the GPT-5.4 Thinking system card if your compliance program tracks safety posture changes per major model release.

FAQ