OpenAI GPT-5.4

GPT-5.4, released on March 5, 2026, is OpenAI's flagship frontier model and the culmination of their mixture-of-experts architecture scaling. It features a 1.05M token context window (922K input + 128K output), the largest usable context among closed-source competitors at launch. The model comes in five variants — Standard, Thinking, Pro, Mini, and Nano — spanning from a $180/month Pro subscription tier down to the lightweight Nano designed for on-device deployment.

The headline benchmark result is a perfect 100% on AIME 2025, the first model to achieve a flawless score on the American Invitational Mathematics Examination. But arguably more significant is its OSWorld performance: 75% versus the human expert baseline of 72.4%, marking the first time an AI model has surpassed human-level performance on real-world computer use tasks including web browsing, file management, and application interaction.

GPT-5.4's architecture uses a mixture-of-experts approach with a knowledge cutoff of August 2025. The Thinking variant introduces extended reasoning with a notable trade-off: time-to-first-token can reach 153 seconds on complex problems as the model works through its chain-of-thought before generating output. The Pro variant ($30 input / $180 output per MTok) unlocks the deepest reasoning capabilities, while Mini ($0.40/$1.60) offers a compelling balance for production workloads.

One practical consideration is cached input pricing at $1.25/MTok, which makes repeated prompts with shared prefixes substantially cheaper. The model scores well across the board — 92.8% GPQA Diamond, 57.7% SWE-bench Pro, and a tied-#1 AA Intelligence Index of 57 — positioning it as a strong generalist that particularly excels at agentic computer use and mathematical reasoning.