Native computer use, a million-token window, and a quiet architectural shift that matters more than any benchmark
On March 5, OpenAI launched GPT-5.4 and two variants — GPT-5.4 Thinking, a reasoning model, and GPT-5.4 Pro — in what amounts to the company's most deliberate pivot from conversational AI to professional execution engine.
The headline capabilities are impressive on their own terms: native computer use (browser and desktop control without third-party scaffolding), a 1-million-token context window in the API and Codex, absorption of GPT-5.3-Codex's coding capabilities, and native ChatGPT integrations for Excel and Google Sheets. The numbers back the ambition — 83% on OpenAI's GDPval knowledge-work benchmark (a record), top scores on OSWorld-Verified and WebArena Verified computer-use benchmarks that reportedly outperform the human baseline, and 33% fewer factual errors per claim versus GPT-5.2.
But the most architecturally interesting feature is the one getting the least attention.
Previous agentic systems faced a scaling problem that rarely made headlines but ate budgets alive. Every tool an agent could potentially use — APIs, database connectors, code interpreters, browser controls — had to be described in the system prompt upfront. As the number of available tools grew, so did the token cost of every single request, regardless of whether most tools were ever invoked.
GPT-5.4's Tool Search inverts that model entirely. Instead of front-loading tool definitions, the model looks up capabilities on demand — like a worker consulting a reference manual rather than memorizing the entire thing before starting a shift.
The efficiency gain scales with complexity. A pipeline with ten tools wastes little on upfront descriptions. A pipeline with two hundred — the kind enterprise deployments are rapidly approaching — wastes enormous amounts. Tool Search turns that linear cost into something closer to constant. For companies building serious agentic infrastructure on top of OpenAI's API, this is likely the feature that matters most.
Alongside the launch, OpenAI released a new chain-of-thought controllability evaluation — the first time the company has publicly scrutinized whether a model's internal reasoning can be steered toward harmful outputs. The GPT-5.4 Thinking System Card, available via OpenAI's Deployment Safety Hub, details the methodology.
This matters more than it might seem. When a model operates as a conversational assistant, the attack surface is relatively constrained: a user types something, a model responds. When a model operates as an autonomous agent controlling a browser and desktop over extended horizons, the attack surface expands dramatically. The question isn't just whether the model can be tricked into saying something harmful — it's whether the model's planning process itself can be corrupted.
OpenAI's evaluation doesn't resolve that question, but it's the first serious public acknowledgment that agentic capability and agentic safety are genuinely in tension. That tension is going to define the next phase of AI development, and GPT-5.4 is where it becomes concrete.
Fortune framed the launch as "a direct shot at Anthropic." That's accurate but incomplete. GPT-5.4's computer use puts it in direct competition with Anthropic's Claude computer use capabilities, yes. But Tool Search is aimed at a different competitor entirely: the open-source agentic frameworks — LangChain, CrewAI, AutoGen — that have been building their own tool-management abstractions. OpenAI is making the case that tool orchestration belongs in the model layer, not the middleware.
Whether that argument holds will depend on whether Tool Search's on-demand lookups are fast enough and accurate enough to replace purpose-built orchestration. Early reports from the API suggest the latency is acceptable. But "acceptable" is a low bar for enterprise systems where every additional millisecond of agent response time compounds across thousands of concurrent sessions.
GPT-5.4 is OpenAI's clearest statement yet that the future of the company is not chatbots — it's professional automation. The model can control your computer, read a million tokens of context, look up its own tools, and reason through multi-step tasks. The benchmarks say it does all of this better than any model before it, and in some cases better than humans.
The question that should keep enterprise buyers up at night isn't whether the model is capable enough. It's whether the safety evaluation framework is maturing at the same pace as the capability. On that score, the chain-of-thought controllability evaluation is a promising start — but it's only a start.
Sources: OpenAI announcement page (openai.com), TechCrunch (March 5), DataCamp technical breakdown, GPT-5.4 Thinking System Card (OpenAI Deployment Safety Hub).