OpenAI's GPT-5.6 Price Cut: What Luna at $0.20 and Terra at $2 Do to Your Margins

Key takeaways

• Luna is now $0.20 input / $1.20 output per 1M tokens, down 80%. Terra is now $2 / $12 per 1M tokens, down 20%.
• Sol pricing is unchanged. Fast mode replaces Priority Processing and runs up to 2.5x faster at twice the price, with no change in intelligence.
• The cuts are asymmetric, so the gap between the cheap tier and the mid tier widened. Routing decisions now matter more than they did last week.
• Output tokens still cost 6x input on both models, so your blended rate depends on your output-to-input ratio, not on the headline price.
• Re-run your unit economics before you re-price anything. A cost cut you do not model is a margin gain you do not capture.

What exactly changed on July 30, 2026?

OpenAI passed serving efficiency gains through to API pricing. GPT-5.6 Luna, the fastest and cheapest model in the family, dropped 80% to $0.20 per 1M input tokens and $1.20 per 1M output tokens. GPT-5.6 Terra, the balanced everyday model, dropped 20% to $2 and $12. Sol is untouched, but Priority Processing is replaced by Fast mode: up to 2.5x faster than standard processing at double the price, same model quality. The lower Luna and Terra rates also flow through to how usage is counted against Codex and ChatGPT Work subscriptions.

Those are the announced numbers. Everything below is what they do to a P&L.

Why does an 80% cut on one model and 20% on another change your architecture?

Because the two tiers just moved apart, not together.

Before the change, choosing between the cheap tier and the mid tier was a quality-versus-cost trade you could reasonably resolve by defaulting to the safer, stronger model. After the change, the ratio between them is wide enough that the default is expensive.

Compare like for like on the new rates. Say a product processes 5M input tokens and 1M output tokens per day:

• On Terra: 5 x $2 + 1 x $12 = $22 per day, roughly $660 per month.
• On Luna: 5 x $0.20 + 1 x $1.20 = $2.20 per day, roughly $66 per month.

That is a 10x difference between tiers, against a 20% discount within a tier. If you spend the week celebrating the Terra cut and none of it re-examining which steps actually need Terra, you left most of the money on the table. This is the observation the announcement does not make for you: the price cut is the smaller lever, and the routing decision it unlocks is the larger one.

How do you work out your real blended rate?

The headline price is two numbers, and your bill is a weighted average of them. On both Luna and Terra, output tokens cost 6x input tokens. So your effective cost per 1M tokens moves with your output-to-input ratio.

A retrieval-heavy workload that stuffs 20k tokens of context in and returns a 300 token answer sits close to the input price. An agent loop that reasons at length and writes code sits close to the output price. Two companies on the same model, at the same list price, can be a multiple apart on effective cost.

Practical steps:

1. 1Pull actual input and output token counts per endpoint from your logs, not estimates.
2. 2Compute cost per request, then cost per active user per month.
3. 3Compare that against your price per seat or per credit to get true gross margin, not blended platform margin.

You can model these scenarios side by side in Calcaas if you would rather not rebuild the spreadsheet.

Should you cut your prices because your costs fell?

Usually not immediately, and not by the same percentage.

Price follows value delivered, cost follows infrastructure. When inference gets cheaper, you have three options: keep the margin, spend it on more capable behaviour per request, or hand it to customers to win on price. Most early-stage products should do the second. A cheaper model per token lets you afford more calls per task, more retries, more verification, better output quality at the same unit cost.

The one thing to avoid is cutting list price reflexively. Price cuts are hard to reverse and your competitors got the same discount on the same day, so it buys less differentiation than it feels like it should.

What about Fast mode?

Fast mode is a latency product, not a cost product. Doubling the price for up to 2.5x speed only pays off where response time changes conversion, retention or an SLA. Model it as a segment: if fast responses belong to one paid tier, price that tier against the 2x inference cost rather than averaging the premium across everyone.

The takeaway: a price cut only reaches your margin if you re-run the model behind it, and the asymmetry between Luna and Terra is worth more than either discount on its own.

You can rebuild your token cost, tier and margin scenarios against the new July 30 rates in [Calcaas](https://calcaas.com).