Migrating: GPT-5.4 to GPT-5.6 Terra

The migration worth checking before the more obvious GPT-5.4-to-Sol move: Terra matches GPT-5.4 on window and rate-limit table, and costs less per token, not more.

What actually moves

The knowledge cutoff jumps forward to the GPT-5.6 generation. Price falls on input, cached input and output at every tier both publish — an unusual direction for a "move to the newer model" migration, and the reason this pairing exists on this site.

The caching correction

This page used to say that free writes on GPT-5.4 become a real per-write cost on Terra, making cache writes "the one cost that moves in the wrong direction." It doesn't move at all. GPT-5.4's writes were never free: they bill at its ordinary input rate, and that rate is exactly Terra's cache-write rate. What does move is everything around the write — Terra's cache reads, uncached input and output are all cheaper — so caching-heavy workloads come out ahead too.

Behavior to set deliberately

GPT-5.4 defaults to none reasoning effort; Terra defaults to medium and also accepts max. Pin the effort you actually want in the request or config, or the migration changes behavior and token spend at the same time.

What stays put

Context window, output ceiling, rate-limit table and the long-context rule.

Before switching

If price was the only reason to stay on GPT-5.4, Terra removes it. GPT-6 Sol is worth a test in parallel: it shares Terra's input, cached-input and cache-write rates, charges less for output, and carries a newer cutoff.