If you read one line: On July 30 OpenAI cut GPT-5.6 Luna about 80% — to $0.20 / $1.20 per million input/output tokens — and Terra about 20%, to $2 / $12. Sol wasn't cut, but "Priority Processing" is now "Fast mode" (up to 2.5× faster, ~2× the price). If you priced a product or picked a model in early July, that decision now rests on numbers that no longer exist. Re-run the math this week.

Three weeks. That's how long the GPT-5.6 launch pricing from July 9 lasted before OpenAI reset the floor underneath it. For a founder, the interesting part isn't the discount — it's the cadence. When frontier prices move 80% in three weeks, "pick the right model once" stops being a strategy. Model choice is now a thing you re-check on a schedule, like your cloud bill.

What actually changed#

Two prices and one tier name:

The stated reason is efficiency, not charity: OpenAI says gains across its models, inference infrastructure, and — notably — the model optimizing its own production and serving code lowered the cost of running GPT-5.6, and it's passing that down. The subtext, which every outlet names, is that enterprise buyers have gotten loud about AI spend, and Gemini's Flash line plus open-weight models like Kimi K3 are pressing from below.

The one line of code to check#

Most of this is a spreadsheet update. One part is not. If your app tags requests for the low-latency tier, you were writing:

client.responses.create(
    model="gpt-5.6-sol",
    service_tier="priority",   # ← the old name
    input=...,
)

service_tier="priority" still works — requests route to Fast mode automatically, so nothing breaks today. But the tier is now named fast, and that's what the docs and dashboards will speak going forward:

    service_tier="fast",       # ← the new name; "priority" still aliases to it

Update it when you touch that code path. And treat Fast mode as a dial, not a default — 2× the price for 2.5× the speed only pays where latency is the product (interactive chat, live coding), not on a background job a user never watches.

What to do this week#

An 80% cut on your cheapest capable tier is the kind of event that quietly changes which architecture is correct. Three moves, in order:

  1. Recompute cost-per-request on your real traffic mix at the new Luna/Terra prices. If Luna carries a meaningful share of your calls, your gross margin just moved — know by how much before a competitor does.
  2. Re-benchmark the cheap tier against your quality bar. The reason to care about an 80% cut is that it makes Luna worth re-testing on tasks you'd previously escalated to Terra or Sol. Run your actual prompts; keep the downgrade only where quality holds. Our tier-by-tier founder guide and the Luna-vs-Gemini-Flash head-to-head are the starting points.
  3. Set a recurring check. The lesson of a three-week price cut is that the frontier reprices faster than most teams revisit their model choice. Put a monthly "re-run the model math" reminder next to the one you already have for your infra bill. More on the discipline in cut your AI bill after the July price drop.

The through-line of this whole frontier price war: capability keeps getting cheaper, and the advantage moves to whoever notices fastest — then spends the savings on the engineering around the model, which is where the real moat now lives. The model that was too expensive for a feature in early July might be the obvious choice by August. Don't let a stale spreadsheet decide your architecture.