GPT-5.4 vs GPT-5.6 Terra

A generation apart in name, but the practical gap is recency and price rather than raw specs — both publish the same window, the same rate-limit table and the same long-context rule.

What's different

Terra's knowledge cutoff runs well ahead of GPT-5.4's. Terra is also cheaper: lower input, cached-input and output rates at Standard, and the same direction at Batch and Flex — the newer model costs less, which is worth noticing because "newer costs more" doesn't hold on this roster.

The caching wrinkle that isn't

This page used to warn that Terra's cache-write charge could erase its price advantage on a write-heavy workload. That warning rested on reading GPT-5.4's writes as free. They aren't: GPT-5.4 adds no surcharge, so a written token bills at GPT-5.4's ordinary input rate — and that rate is exactly Terra's cache-write rate. Writes cost the same on both models; reads, uncached input and output are all cheaper on Terra. There is no caching pattern under which GPT-5.4 comes out ahead.

Reasoning defaults differ

GPT-5.4 defaults to none — an unspecified request does no reasoning — while Terra defaults to medium and also accepts max. Migrating without setting effort explicitly changes behavior and token use at once.

How to decide

On price and recency Terra wins outright. The reason to keep GPT-5.4 is a validated pipeline you don't want to re-test. If you sign in to Codex with ChatGPT, the choice was made for you: GPT-5.4 left Codex there on August 31, 2026.