GPT-5.6 Terra vs GPT-5.6 Luna

Terra and Luna sit next to each other on price within the GPT-5.6 family, but they diverge on something a price table alone won't show: which rate-limit table they belong to.

The window is the same, the throughput isn't

Both publish the identical context window, output ceiling and knowledge cutoff. Where they split is rate limits — Luna's table allows the same requests per minute as Terra's at Build and Launch and more at Grow, larger token ceilings at every paid tier, and a larger batch queue at Build and Launch (the two match at Grow). The cheaper model has the more generous throughput, which runs against the usual assumption that price and capacity move together.

Price moves the expected way

Luna is cheaper than Terra on every cell — input, cached input, cache writes and output, short and long context, every service tier. That part is the ordinary "smaller model, smaller bill" pattern.

The GPT-6 versions of this choice

Both have newer counterparts at lower prices. GPT-6 Luna undercuts GPT-5.6 Luna everywhere on the same rate-limit table, and GPT-6.1 Terra — the newer Terra, released October 26, 2026 — undercuts GPT-5.6 Terra on every Standard price cell on the same table, with cheaper cached reads. If you're choosing between these two today, put the newer pair on the same shortlist; GPT-5.6 Terra to GPT-6.1 Terra covers that move.

How to decide

For a high-volume, throughput-sensitive workload, Luna's lower price and higher ceiling point the same way. For a task where Luna's answers genuinely fall short, Terra is the step up to try before reaching for a Sol model — run the same evaluation set through both rather than assuming the dearer one is needed.