OpenAI launches GPT-5.4 with native computer use and 1M context

Running heavy agentic workflows usually runs into two infuriating walls. First, the model cannot physically touch your desktop software. Second, when its chain of thought drifts into left field, you have to watch it burn tokens before starting entirely over.

OpenAI tackled both pain points on March 5 with the release of GPT-5.4 and its heavier sibling, GPT-5.4 Pro. Dropping across ChatGPT, the API, and Codex, the new frontier model pushes context length to 1 million tokens and bakes native computer-use capabilities straight into the architecture.

Handing the mouse and keyboard to AI

GPT-5.4 is built to stop talking about tasks and start clicking through them.

On the desktop navigation benchmark OSWorld-Verified, the model scored a 75.0% success rate driving mouse and keyboard actions from raw screenshots. That crushes GPT-5.2 at 47.3% and edges out the human baseline of 72.4%.

It put up similarly high marks across browser environments, clocking 67.3% on WebArena-Verified and 92.8% on Online-Mind2Web using visual inputs alone. To make sure it does not misclick tiny buttons, OpenAI added original-resolution vision support handling up to 10.24 million pixels.

Developers can also set granular confirmation policies so the model pauses for human approval before triggering sensitive actions. Over in Codex, a new interactive Playwright skill lets the model visually test web and Electron apps as it builds them.

Dynamic tool search and mid-flight steering

Scaling up complex jobs usually wrecks your API budget, so OpenAI reworked how tools get loaded.

Instead of dumping dozens of function definitions upfront into every prompt, GPT-5.4 uses tool search. It holds a lightweight catalog and only pulls full tool schemas when it actually decides to invoke them.

On Scale's MCP Atlas evaluation across 36 MCP servers, this dynamic lookup slashed token usage by 47% without hurting task accuracy.

For everyday ChatGPT users running GPT-5.4 Thinking, the killer feature is mid-response steering. The model lays out an upfront plan for complex requests. If you see it going off the rails, you can tweak instructions midway through its reasoning without restarting the chat.

Combined with a 33% reduction in factual false claims compared to GPT-5.2 and an 87.3% score on junior banking spreadsheet modeling, it feels significantly less brittle in corporate settings.

Why it matters

We see GPT-5.4 as OpenAI planting its flag firmly on Anthropic's home turf of agentic desktop interaction.

For developers and knowledge workers, this meaningfully shifts the barrier to real automation. The combination of a 1M context window, high-res screen parsing, and Excel integration turns AI from an advisory chatbot into an operator that can actually finish multi-app knowledge work.

Tool search and mid-response adjustments are the smartest parts of this release. They address real operational pain: token bills ballooning from massive tool schemas and user frustration with runaway reasoning chains.

Across the industry, desktop agency is no longer an experimental research preview; beating human averages on OSWorld sets a brand new competitive bar for frontier models.

Yet the catch remains consistency. Crossing the 75% threshold is impressive on paper, but a one-in-four failure rate is still far too high for unsupervised operations on sensitive production systems. Custom confirmation policies help keep a leash on rogue actions, but if humans have to babysit every click, the speed benefits evaporate. Closing that remaining reliability gap is where the real enterprise war will be fought.