GPT-5.6’s new economics depend on context discipline

OpenAI’s July 30 update changes the operating economics around GPT-5.6 rather than presenting a new capability claim. The company says Luna now costs 80% less and Terra 20% less, while Sol gains an API Fast mode that it says can run up to 2.5 times faster than Standard processing at twice the price. OpenAI also says Fast mode replaces Priority Processing and that existing requests marked as priority remain compatible. For teams already using the family, this is principally a routing and latency decision: lower-tier jobs can become cheaper, while urgent Sol calls carry a clear price premium.

The update follows OpenAI’s June description of the three-model lineup, in which Sol was the flagship, Terra the everyday balanced option, and Luna the fast, lower-cost option. OpenAI had characterized Terra as competitive with GPT-5.5 while being twice as cheap during that earlier preview. The current list prices are $2 per million input tokens and $12 per million output tokens for Terra, versus $0.20 and $1.20 for Luna; Sol pricing is unchanged. That makes the family’s differentiation more explicit, but it does not establish that one tier will outperform another on every workload.

OpenAI’s ARC-AGI-3 write-up is a useful warning against treating the price changes as a simple model leaderboard story. In its own experiment, the company says retaining reasoning and enabling compaction raised Sol’s score on the public task set from 13.3% to 38.3% and reduced output tokens sixfold. OpenAI attributes much of that movement to the harness: the original setup discarded private reasoning and progressively removed old context. The practical inference is that evaluations and production costs depend on orchestration choices, so the reported result should be read as a configuration finding from OpenAI, not an independent measurement of a universally higher model capability.

Amazon’s Bedrock guidance gives that orchestration issue a concrete cost boundary. Its July 30 post says GPT-5.6 prompt-cache reads receive a 90% discount, cache writes cost 1.25 times the uncached input rate, and a workload breaks even once cache reads account for roughly 20% of tokens passing through the cache. Bedrock’s explicit mode lets an implementer place a cache boundary after stable instructions or references, while its implicit mode is the default. For agent loops, retrieval systems, and repeated document-analysis jobs, teams should inspect cached-token and cache-write fields in production instead of assuming a headline price reduction will automatically lower their bill.