Same model. Same Thursday. An 80% cut, landing exactly three weeks after GPT-5.6 launched.
Luna's input price fell from $1 to 20 cents. Output dropped from $6 to $1.20. Combined, $7 became $1.40.
Here is the part that should keep a CFO up at night: the price didn't just drop. It dropped between the moment you planned and the moment you shipped.
The price didn't just drop. It dropped between the moment you planned and the moment you shipped.
If your budget cycle is a quarter and your vendor's repricing cycle is three weeks, the rate you negotiated is already a historical artifact by the time the invoice arrives.
Last week, in the Per-Token Paradox, we argued that usage-based pricing was making bills unpredictable even as per-token rates fell. That was a projection. This week OpenAI made it the event.
The trap isn't that Luna got cheaper. Cheaper is good. The trap is that "the price of Luna" is now a number that moves 80% inside a single planning horizon — which means it isn't a planning primitive anymore.
And the single cut is the small story. The big story is what the price of a model has become.
Pull up OpenAI's own pricing page and you don't find a number. You find a matrix.
Every flagship model now has a short-context column and a long-context column, each with its own input, cached input, cache-writes, and output rate. Luna's short-context input is $0.20; its long-context input is $0.40 — double, for the same model, depending on how much you stuff into the window.
Then there are tabs: Standard, Batch, Flex, and Fast mode. Batch runs cheaper on a delay. Fast mode runs pricier for speed.
Sol Fast mode charges twice the Standard price — $10 input, $60 output — for up to 2.5x the throughput and, in OpenAI's own words, "no change in intelligence." You are now buying latency as a separate SKU from intelligence.
And Priority Processing was renamed Fast mode the same day, July 30, with old requests tagged "priority" automatically routed to the new tier. The label under your line item changed while you weren't looking.
So "what does GPT-5.6 cost" has no answer. It has a lookup table with at least four axes: which model, which context length, which processing tier, and whether the input was cached.
Zoom out past OpenAI and the axes multiply. Some providers reportedly bill batch jobs at roughly one-tenth of burst rates. Google reportedly prices long documents on floating-point operations — actual GPU work — instead of tokens, which can beat per-token billing outright above 100,000 tokens. DeepSeek reportedly discounts off-peak hours, but only for batch.
The per-token standard that OpenAI reportedly introduced in 2022 has fractured into subscriptions, compute bundles, context-aware microtransactions, and prepaid wallets. There is no longer a "rate" to negotiate. There is a surface.
This is the reframe procurement teams keep missing. You cannot negotiate a rate card against a variable that no longer exists as a single number.
You cannot budget against a price that fell 80% mid-quarter. And you especially cannot sign a multi-quarter token commitment to a vendor that is repricing under IPO pressure.
That last point is not speculation. OpenAI cut these prices while facing Moonshot's Kimi K3, Z.ai's GLM-5.2, and Google's Gemini 3.6 Flash — all while analysts note the cuts thin the very margins IPO investors will scrutinize.
A vendor caught between winning cost-sensitive customers and proving profitability to future shareholders will keep moving the price. Committing to a fixed rate against that is, arguably, committing to be wrong.
Now connect this to last issue's Uber blowup. That was the postpaid symptom: a bill that arrived after the spend, too late to stop it.
This week's 80% cut is the same disease at the planning end. Postpaid is "you find out after it happened, on an invoice, once it's too late to stop it." A per-token rate card is the same failure moved upstream — you commit to a number before it happens, then watch it drift 80% away from reality.
Both mistakes share a root cause: treating the price as fixed and knowable. It is neither.
So what do you actually do? Stop committing to prices. Start committing to spend ceilings and routing policies.
First, treat every model as interchangeable compute behind a router. OpenAI is practically instructing you to do this — their own coding-workflow example uses Sol to plan and Luna to implement, switching models mid-task.
If Luna can drop 80% overnight and land on the price-performance frontier, and if Terra reportedly now does work that recently required a top-tier model, then hardcoding a model name is hardcoding a rate that will be stale next Thursday.
Define policies, not names: "use the cheapest model that scores above our accuracy threshold." Let the router re-shop the market on every request. When a vendor cuts 80%, you capture it automatically instead of scheduling a migration.
Second, replace rate negotiation with wallet-enforced ceilings. Postpaid means unbounded spend unless you're watching. Prepaid means the maximum possible spend for a period is the number you topped up — full stop, no month-end surprise.
Top up ₹5,000 and ₹5,000 is the ceiling. There is no credit extended, only balance drawn down in real time. That is a budget you can actually defend, because it's enforced by the system rather than discovered on an invoice.
Third, push the ceiling down to the key. Per-key caps let one account-level wallet split into engineering at ₹2,000, data at ₹1,500, and a sandbox at ₹500 — independently capped from the same pool.
When an experimental script in the sandbox blows through ₹500 — the runaway retry loop, the agent calling the same endpoint hundreds of times a minute — it stops there. It cannot touch the other four keys. That is the containment postpaid billing structurally cannot give you.
Notice what all three moves have in common. None of them require you to know next quarter's price. They require you to know your ceiling and your quality bar — the two things that are actually yours to control.
The old procurement instinct was to lock in a rate and feel safe. In a market where a frontier model's combined price collapses from $7 to $1.40 in three weeks, that instinct is now the risk.
The rate is not a number you own. It is a number