OpenAI ships GPT-5.4 with native computer use and a 1M-token API context (March 2026)

On 5 March 2026 OpenAI released GPT-5.4, positioning it as its most capable and efficient model for professional work. The launch came in three forms: a standard model, a GPT-5.4 Thinking reasoning variant inside ChatGPT, and GPT-5.4 Pro for maximum performance. The model rolled out the same day across ChatGPT, the API (as gpt-5.4) and Codex, and OpenAI announced that the older GPT-5.2 Thinking model would be retired on 5 June 2026.

The headline capabilities are agentic. GPT-5.4 has native computer use, meaning it can operate websites and desktop software from screenshots plus keyboard and mouse actions, and OpenAI reported a score of 75.0% on OSWorld-Verified, above the 72.4% it attributes to human testers. On its own GDPval benchmark for knowledge-work deliverables it scored 83.0%, and it posted 57.7% on SWE-Bench Pro and 82.7% on BrowseComp (89.3% for the Pro variant). The API supports a 1M-token context window, experimentally in Codex, while ChatGPT keeps a 272K window; requests beyond 272K are billed at double rate. A new Tool Search mechanism lets the model fetch tool definitions on demand rather than loading them all up front, which OpenAI says cuts token use by 47% in tool-heavy workflows.

Pricing is $2.50 per million input tokens and $15 per million output tokens for GPT-5.4, and $30/$180 for GPT-5.4 Pro, with batch and flex tiers at half price and a priority tier at double. OpenAI also reported accuracy gains versus GPT-5.2: individual claims are 33% less likely to be false and whole responses 18% less likely to contain any error.

The safety notes are as relevant to enterprises as the benchmarks. OpenAI kept the high cyber-capability classification it had applied to GPT-5.3-Codex, and paired the release with expanded cyber safety systems, monitoring, trusted-access controls and blocking of higher-risk requests on zero-data-retention surfaces. It also published research on chain-of-thought controllability, concluding that GPT-5.4 Thinking shows a low ability to obscure its reasoning, which the company frames as evidence that CoT monitoring remains a workable oversight tool. The company acknowledged that its detection systems will generate some false positives.

Compared with the incremental GPT-5.x updates of late 2025, this release marks the point where computer operation, long context and tool orchestration are bundled into a general-purpose flagship rather than sold as separate agent products. That closes the gap with Anthropic's and Google's agentic offerings and makes autonomous desktop and browser agents a default option for any organisation on the OpenAI stack.

What it means for leaders

• Treat computer-use agents as a new privileged identity. An agent that can click through your SaaS estate needs scoped credentials, session recording and kill switches, not an employee's shared login. Map this against OWASP LLM Top 10 excessive-agency guidance.
• Re-baseline data flows before enabling 1M-token contexts. Very large windows make it trivial to paste entire repositories or customer datasets into a prompt; update DLP rules and retention settings accordingly.
• Budget for the tiering. The 2x charge above 272K tokens and the Pro pricing at $180 per million output tokens can surprise teams migrating long-document pipelines; put cost telemetry in place before rollout.
• Log the cyber-risk classification in your AI inventory. Under ISO/IEC 42001 and NIST AI RMF (Map and Govern functions), a vendor's own high-capability rating is material to your risk assessment and to any EU AI Act GPAI due diligence.
• Plan the GPT-5.2 Thinking retirement now. Three months is short for regression testing prompts and evaluation suites in regulated workflows.