GPT-5.4 was announced on March 5. Broad availability came the following week. That gap matters for practitioners who held off, the model is now accessible at production scale, and the capability questions are no longer theoretical.
OpenAI states the model carries a one-million-token context window and an “extreme reasoning mode” built for tasks requiring multiple hours and high reliability. A separate “Thinking mode” allows users to interrupt the model during generation and steer it in a new direction, a workflow change that meaningfully alters how prompt engineers approach long-running tasks, according to devFlokers’ March 13 model release roundup.
The model’s most technically significant addition is native computer-use. According to The Next Web’s coverage, GPT-5.4 introduces native computer-use that allows it to interact with desktop environments directly. OpenAI reports the model achieved a 75% success rate on the OSWorld-Verified benchmark, a self-reported figure, not yet independently reproduced. Secondary sources, including TechStrong AI, cite a human baseline of 72.4% on the same benchmark.
Three caveats apply to these numbers. OSWorld-Verified is a benchmark designed to test AI agents on real desktop tasks. The 75% figure comes from OpenAI’s own reported results. No independent third-party reproduction of that score has been confirmed in current reporting. Practitioners should treat it as a directional signal, not a verified ceiling.
GPT-5.4 isn’t the only frontier model that landed this week. xAI’s Grok 4.20 and Google’s Gemini 3.1 Pro are both reported as simultaneous releases, each with different headline claims. Deeper comparative coverage is coming once additional verification on those models is complete.
The practical question for AI development teams right now is straightforward: computer-use at or above human baseline performance on standardized tasks changes the build calculus for agentic workflows. The 1M token context opens document-scale reasoning that wasn’t previously viable in a single API call. Whether those vendor-stated numbers hold in production workloads is the verification question every practitioner should run before committing infrastructure to them.