GPT-5.6 price cuts make model routing a more explicit engineering choice

OpenAI says GPT-5.6 Luna now costs 80% less and Terra 20% less, while Sol gains an API Fast mode that delivers up to 2.5 times Standard speed at twice the price. The company says Fast replaces Priority Processing and that existing API requests marked `priority` will continue to work. This is a pricing and serving-policy change, not a new release claim, but it materially changes the cost model for teams already using the GPT-5.6 family.

The published API rates are $0.20 per million input tokens and $1.20 per million output tokens for Luna, versus $2 and $12 for Terra; Sol pricing is unchanged. OpenAI also says Terra and Luna usage will consume fewer credits in Codex and ChatGPT Work even though subscription prices and quota budgets remain unchanged. That means teams should separate a lower metered API bill from any changes in product-plan limits, because the two are described as different mechanisms.

The available documentation frames the three models as distinct operating tiers rather than interchangeable names. AWS describes Sol as suited to autonomous coding, security research, scientific analysis, and deep multi-step reasoning; Terra as the balanced production option; and Luna as the high-volume, latency-sensitive option. AWS also lists a shared 272K-token context window and Responses API support, so the routing choice can increasingly be driven by task risk, latency, and cost rather than by a different integration surface.

Prompt caching is an important qualifier to the headline price reduction for Bedrock users. AWS says GPT-5.6 supports implicit caching by default and an explicit mode that lets callers choose the reusable prefix. Cache reads receive a 90% discount, writes cost 1.25 times the normal input rate, and AWS estimates that a workload begins lowering net input cost when cache reads are about 20% of tokens flowing through the cache. Repeated-agent-context workloads therefore need cache-hit measurements, not just a comparison of list prices.

For a practical rollout, use Sol where a higher reasoning budget or faster completion changes the outcome, then evaluate Terra or Luna for bounded implementation, classification, summarization, routing, and tool-loop steps. AWS advises teams moving from GPT-5.5 or GPT-5.4 to begin at their current reasoning effort and test one level lower, while OpenAI gives the same general case for assigning more intelligence only where it improves the result. That is a testable migration plan, but not a guarantee that a lower tier will preserve quality for a particular workload.

The strongest performance comparisons in the announcement are OpenAI's own measurements and customer statements, so they should be treated as provider-reported evidence rather than an independent ranking. The verified facts are the advertised price, Fast-mode, API, and Bedrock caching changes; the outcome for a production system remains uncertain until teams measure quality, token use, cache behavior, regional availability, and end-to-end latency on their own tasks. In particular, Fast mode's speed premium and cache-write cost can erase the apparent savings for the wrong traffic pattern.