OpenAI Launches GPT-5.4 with Native Computer Control and Financial Integration

OpenAI officially announced the release of its advanced general-purpose model, GPT-5.4, in early March 2026, signaling a strategic shift toward agentic workflows and direct software operation. This new iteration unifies sophisticated reasoning, advanced coding capabilities inherited from GPT-5.3-Codex, and novel computer-use functionality across its GPT-5.4 Thinking and GPT-5.4 Pro variants.

A core advancement is the model's native ability to operate computers, interpreting screenshots and issuing virtual mouse and keyboard commands via libraries such as Playwright. This capability was demonstrated on the OSWorld-Verified benchmark, where GPT-5.4 achieved a 75.0% success rate, surpassing the listed human baseline of 72.4% and significantly outperforming GPT-5.2's 47.3% rate. The model's improved reliability is quantified by data indicating that individual factual claims are 33% less likely to be false than those from GPT-5.2, with full responses containing 18% fewer overall errors.

The launch is strategically focused on high-value enterprise applications, particularly through deep integration into financial software ecosystems. This includes the beta release of ChatGPT for Excel, embedded as a sidebar add-in. This integration enables professionals to construct, audit, and update complex financial models using natural language, with the AI linking actions directly to specific cells and requesting user confirmation before executing changes.

Furthermore, GPT-5.4 is designed for crucial institutional data integrations, facilitating the incorporation of market data, company filings, and research reports into valuation and underwriting workflows. For developers, the API now supports an extensive context window of up to 1 million tokens, enabling the ingestion of entire codebases or lengthy documentation, although double billing applies above 272,000 input tokens. To manage computational scale, OpenAI introduced a Tool Search feature in the API, which, in testing, reduced total token usage by around 47% on complex tool-use tasks.

The model's performance across general knowledge work is highlighted by an 83.0% success rate on the GDPval benchmark across 44 occupations, exceeding GPT-5.2's 71.0%. This introduction intensifies competition with rivals like Anthropic, particularly in the finance sector. The GPT-5.4 Thinking API pricing is set at $2.50 per million input tokens and $15 per million output tokens, while the GPT-5.4 Pro variant is priced higher. OpenAI suggests that the model's improved token efficiency will offset these per-token rate increases for many complex deployments, and the preceding GPT-5.2 Thinking model is expected to be phased out over the course of 2026.