What happened

OpenAI cut prices for two of its three GPT-5.6 models on July 30. GPT-5.6 Luna, the cheapest and fastest tier, now costs 80% less to run through the API. The API is the interface developers use to call OpenAI’s models directly from their own code, instead of through the ChatGPT app. Terra, the everyday tier, costs 20% less.

OpenAI introduced the three-tier GPT-5.6 family three weeks ago: Sol as the top-reasoning model, Terra as the everyday default, and Luna as the fast, cheap option. This update doesn’t add new capabilities. It makes two of the three tiers cheaper and adds a new paid option that makes the third, Sol, faster.

What it actually costs now

Per OpenAI’s published API pricing, Luna now costs $0.20 per million input tokens and $1.20 per million output tokens. A token is roughly a word fragment, so a million tokens is a rough stand-in for word count, not an exact one. OpenAI says that’s an 80% cut. Work backward from that, and Luna’s old price comes out to about $1.00 per million input tokens and $6.00 per million output tokens.

Terra now costs $2 per million input tokens and $12 per million output tokens, down roughly 20% from about $2.50 and $15.

Reused context, like the conversation history a coding agent keeps referencing across a long session, is billed differently. It gets a lower “cached input” rate: $0.02 per million tokens for Luna, $0.20 for Terra. That’s the price for tokens the model has already processed once and doesn’t have to read again from scratch.

A paid way to make Sol faster

OpenAI also replaced its Priority Processing option with something it calls Fast mode. For GPT-5.6 Sol, the top-reasoning model, Fast mode runs up to 2.5 times faster than standard processing, at twice the price. OpenAI says the model’s actual intelligence doesn’t change: you’re paying only for speed.

That’s a real tradeoff, not a free upgrade. Say your app has a user actively waiting on a response. Doubling the cost for a faster answer might be worth it there. For background work nobody is watching in real time, standard processing at the lower price is still the better default.

Who should care

If your own code or tool calls the OpenAI API directly, for example a custom coding agent you built, these prices apply immediately. Your bill drops as long as you stay off Fast mode.

Codex (OpenAI’s coding agent) and ChatGPT Work (OpenAI’s team subscription plan) work differently. OpenAI says these lower backend costs are reflected in how your usage allowance gets counted. That means the same monthly allowance should now stretch across more requests.

If you route tasks between models to control cost, Luna’s 80% cut changes the math. OpenAI positions Luna as able to use tools and complete multi-step workflows, not just answer single questions. That makes the lower price relevant for more than throwaway tasks.

What builders should do next

Don’t assume your bill dropped by the same percentage as the sticker price. Your mix of input and output tokens affects the real number. Pick one task you currently run on Terra that could tolerate a lighter model. Run it once on Terra and once on Luna with an identical prompt.

Compare four things: did it pass on the first try, how many corrections it needed before it was usable, how long each run took, and what it actually cost at the new prices. If Luna holds up on quality for that task, the 80% price cut makes it worth defaulting to there instead of Terra.

If you have a task where a user waits on the response in real time, test standard Sol against Fast mode on that same task. Time each run with a stopwatch and compare the cost side by side. Only keep paying double if the speed difference is one your users would actually notice.

End of article