OpenAI released GPT-5.4 mini and nano today. The mini sits at $0.75 per million input tokens, the nano at $0.20. The full GPT-5.4, announced ten days ago, costs substantially more for both input and output.

The interesting number is 54.4%. That's GPT-5.4 mini on SWE-Bench Pro, the benchmark that tests professional-grade coding on real repositories. The full GPT-5.4 scores 57.7%. Three percentage points separate the cheap model from the expensive one on the hardest coding evaluation OpenAI publishes. Context window drops from 1.05 million tokens to 400,000. Still enormous.

OpenAI frames this as a multi-model architecture play: the flagship plans, the mini executes. That's a reasonable pitch for agentic workflows where you're orchestrating dozens of parallel calls and per-token cost actually matters. GitHub Copilot already ships it at a 0.33x premium request multiplier, which tells you where the volume is heading.

The pattern repeats across every model family now. The mid-tier eats the flagship, the mini eats the mid-tier, and within six months the nano handles tasks that needed the flagship a year ago. The real product isn't any single model. It's the pricing curve.

Sources: