GPT-5.4 Review

GPT-5.4 arrives with two claims and largely honors both. The first is accuracy: it invents fewer facts than the model it replaces by a wide margin, and testers running research tasks report answers they no longer feel obliged to check line by line. The second is computer use. It reads a screen, moves a pointer and works a keyboard well enough to clear a benchmark of ordinary desktop chores above the average human score, which turns a whole category of tedious work into something that can be handed over rather than described.

Deep research is quicker and spends fewer tokens getting there. The complaints are just as consistent. It drifts from instructions on longer prompts, doing something adjacent to what was asked and doing it with total confidence, and the smaller variants cost a multiple of what the earlier small models did, with the rate stepping up again once a prompt runs long, which weakens the case for putting it anywhere near a high volume pipeline.

As the model to reach for on hard agentic work it is the strongest general option on sale. As the one you call a million times a day, it is expensive enough to think twice about.