OpenAI just dropped GPT-5.4, and two things stand out: it can use your computer the way you do, and it can build Excel financial models without touching a formula. Both are bigger deals than they sound.
Computer use that actually works
The headline capability is native computer use. Previous AI models could theoretically interact with computers through APIs and tools, but GPT-5.4 is the first general-purpose model shipped with this as a core feature. It reads screenshots, issues mouse and keyboard commands, and can browse the web persistently to find information.
The benchmark numbers tell the story. On OSWorld-Verified, which tests desktop navigation using screenshots and actions, GPT-5.4 hits 75% success. That beats reported human performance at 72.4%. On BrowseComp, which measures web research, GPT-5.4 Pro reaches 89.3%. These are not toy benchmarks. They are testing whether an AI can do the kind of multi-step, persistent work that actual knowledge workers do.
OpenAI calls this agentic behavior. The model can keep state across many actions rather than just responding once. That shift from chatbot to agent is what the entire industry has been racing toward, and GPT-5.4 represents a concrete step forward.
Excel and Sheets, finally useful
The second big piece is the spreadsheet integration. OpenAI released ChatGPT for Excel and Google Sheets in beta. It embeds directly into spreadsheets to build, analyze, and update financial models using the formulas and structures teams already use.
This is different from previous AI spreadsheet features, which mostly generated a formula or two. GPT-5.4 can handle complex multi-sheet financial models, scenario analysis, and recurring workflows like earnings previews or DCF analysis. OpenAI claims an internal investment banking benchmark jumped from 43.7% accuracy with GPT-5 to 88.0% with GPT-5.4.
The timing matters. Anthropic already rolled out Claude for finance with Excel integration. Now OpenAI is responding with its own suite, including integrations with FactSet, MSCI, Third Bridge, and Moody’s. The finance sector is becoming a battleground.
Efficiency gains, not just capability gains
OpenAI emphasized that GPT-5.4 uses far fewer tokens. On some tasks, it uses 47% fewer tokens than its predecessors. The model also introduces tool search in the API, which retrieves full tool definitions only when needed rather than dumping everything into the prompt. On a benchmark with 36 MCP servers enabled, this reduced token usage by 47% while maintaining the same accuracy.
The context window supports up to 1 million tokens, allowing agents to plan, execute, and verify tasks across long horizons. That matters for the kind of multi-step workflows that computer use enables.
The pricing picture
Here is the catch: GPT-5.4 is expensive. The API pricing is 15 per million output tokens. GPT-5.4 Pro runs 180 respectively. That places it among the most expensive models available. Compared to alternatives like Gemini 3.1 Flash-Lite at 0.25, the cost difference is significant.
For enterprises that need the capability, the price may be justified. But for many use cases, cheaper models will suffice. OpenAI is betting that the efficiency gains and capability improvements justify the premium.
What this means
GPT-5.4 signals two shifts. First, the agent era is here. Computer use is no longer a research prototype or a niche feature. It is shipping in a production model and can handle real workloads. Second, AI is moving into the daily tools where white-collar work happens. Excel and Sheets are just the start. The integration with financial data providers suggests a broader strategy to embed AI where business decisions get made.
For developers, the tool search feature in the API is worth exploring. It addresses a real pain point: as tool ecosystems grow, passing every tool definition with every request becomes expensive and slow. Retrieving only what is needed is a structural fix.
For enterprises, the spreadsheet integrations are the headline. If your team spends hours on financial models, GPT-5.4 can accelerate that workflow. Just be ready for the price tag.