OpenAI pitches GPT-6 Sol and Luna as up to 50% cheaper than comparable GPT-5.6 models

AWS's wording is narrower than OpenAI's ^[5]. Sol "is built for the demanding, recurring work of development and operations," its weekly roundup said, while Luna "makes focused, repeatable tasks practical at high volume" ^[8]. The roundup's author wrote that the aim is to "match the model to the job instead of reaching for the biggest one every time" ^[9]. I think sorting by job is the right design for the two models under Astra ^[3]. Routing code knows the request type at call time, and the two names now map to request types.

OpenAI's figure, as the dev.to write-up reports it, is up to 50% off comparable GPT-5.6 options ^[1]. That write-up is the most detailed account, and it ends with a pitch for an AI visibility scanner ^[10]; I did not go there expecting a rate card. It does note that the announcement material lacks a complete price card or a savings figure that holds for every use case ^[15]. AWS says only that both models ship at "significantly lower pricing than their GPT-5.6 predecessors" ^[8].

So the 50% is a ceiling. For it to reach a real invoice, the GPT-5.6 model a team runs today has to be the one OpenAI measured against. The new model also has to finish the job in roughly the same tokens and attempts. If cost scales with tokens, a 50% cut lets a job use up to twice the tokens before it costs more than before ^[14]. Retries and fallbacks from Luna to Sol spend that margin first.

With a separate identifier for each model ^[11], moving a call is a one-string change. The work is deciding which calls move. The dev.to piece says teams should validate quality, reliability and cost against their own workload before standardizing on a model ^[7]. In my context I would put Luna on the repetitive, well-specified calls with a fixed output shape first, where a regression shows up quickly. Sol keeps the multi-step jobs until the bills come back.

Agent jobs add a second price variable. OpenAI describes caching and session-management changes that support agent-based workflows, alongside gains in factuality, coding and computer use ^[6]. An agent that resends its context on each turn pays for that context again and again. For those jobs I'd expect the cache terms to move the bill as much as the per-token rate.