GPT-5.6 price-performance analysis: Luna drops 80%

The useful part of a model price cut is rarely the press line. It is the spreadsheet that appears two hours later, with one tab for latency, one for output tokens, and one for the procurement person asking why last quarter's architecture still exists. OpenAI's GPT-5.6 update is one of those moments: less a single discount than a reminder that model choice is becoming workload accounting with nicer branding.

What OpenAI changed,

according to OpenAI OpenAI said on July 30, 2026, that GPT-5.6 Luna will cost "80% less," while GPT-5.6 Terra will cost "20% less," according to its announcement, Advancing the price-performance frontier with GPT-5.6. The company described Luna as its fastest and most affordable model, Terra as its balanced model for everyday work, and Sol as getting faster performance in the API. That is the practical split: high volume work moves toward Luna, routine enterprise work toward Terra, and time sensitive or higher performance API workloads toward Sol. ETIH EdTech News reported that OpenAI has begun a limited preview of GPT-5.6 with Sol, Terra, and Luna, with broader access through ChatGPT, Codex, and the API planned in the coming weeks. It also described Sol as the flagship model, Terra as positioned for general workloads, and Luna as designed around speed and cost. Put less ceremonially, OpenAI is telling builders not to ask which model is best. It is asking them to prove which model is best for the invoice.

Who is affected,

according to ETIH EdTech News Developers, businesses, and users of OpenAI's AI products are the immediate audience, according to ETIH EdTech News. The affected systems are not only new prototypes, but any product where a model call happens often enough that token costs become a product constraint. That includes coding assistants, support automation, internal knowledge tools, workflow agents, and classroom or training products that need predictable usage costs. The governance point is boring and therefore important: a cheaper model does not remove the need to document model selection. If a team routes learner support, employee workflows, or customer decisions through GPT-5.6 Luna because the price changed, it should record why Luna is sufficient for that workload. If it keeps Sol for higher stakes tasks, it should record the same reasoning. Auditors rarely enjoy vibes as evidence.

What changes in practice,

according to OpenAI OpenAI said the lower Luna and Terra prices are also reflected in how usage is counted against paid subscriptions when using Codex and ChatGPT Work. That matters because many teams do not experience model cost only through raw API billing. They experience it through subscription limits, internal chargebacks, and the quiet politics of which department is allowed to run the expensive workflow. The practical move is to stop treating the model picker as a personality test. Route high volume, lower risk, repetitive work to the model OpenAI is pricing for scale. Keep Sol where API speed or capability changes the business outcome. Put Terra in the middle where general workloads need a balance of cost and performance, which is exactly how OpenAI is positioning it.

The benchmark caveat,

according to Artificial Analysis Artificial Analysis ranks GPT-5.6 Sol max as its most intelligent OpenAI model with an intelligence index score of 59, while GPT-5.6 Luna max appears in its output speed ranking at 195 t/s. Those numbers are useful, but they do not settle your architecture. A benchmark can tell you where to start testing. It cannot tell you whether your support bot fails on refunds, whether your coding agent burns output tokens, or whether your school district procurement office will accept the risk note. This is where the compliance file earns its keep. Record the workload, the model tier, the expected volume, the fallback model, and the reason the chosen model is adequate. If the price cut pushes a migration, run the old and new model against the cases that actually matter. LinkedIn may call that overthinking. Your incident review will call it Tuesday. The next thing to watch is not only whether competitors answer with their own price moves. Watch whether model menus keep hardening into fast, balanced, and high performance lanes, because that is where product architecture, vendor negotiations, and AI governance paperwork will meet. Builders who can measure workload value per model call will have better budgets and fewer awkward meetings.