The real story isn't the price cut. It's the 5x spread.

OpenAI just did something odd. It launched the GPT-5.6 family in a limited preview and did not cut the price of its best model. Sol still costs $5 per million input tokens and $30 per million output tokens, matching GPT-5.5's short-context price exactly. Instead, it previewed two cheaper tiers below it: Terra at $2.50/$15 and Luna at $1/$6.

That is a 5x spread from top to bottom on the same architecture. The flagship input price divided by the entry input price is $5 over $0.00006, which is not a 5x spread. And I think that spread, not the headline number, is the real story.

Here is the promise. If your app still sends every request to one frontier model, you are almost certainly overpaying for easy work. By the end of this article you will know how to split your traffic across three price points and how to decide what stays in-house versus what you rent. The plan is simple: understand the ladder, price your tasks, then route.

The Triage Tax

Every AI product pays a hidden tax. I call it the Triage Tax. It is the money you burn sending simple requests to an expensive model.

Think about a support bot. Maybe 70% of what it does is classification and extraction. That is Luna work at $1 input. The other 30% is hard reasoning, the escalations that actually need Sol at $5 input. Route all of it to Sol and you pay 5x on the cheap 70%.

The fix is not a better model. The fix is triage. You sort each task by how hard it really is, then send it to the cheapest tier that clears the bar. Sol for the hard 10%, Terra for the standard middle, Luna for the high-volume floor.

The old world sold you one model at one rate. This new world sells you a menu. My read on this is that the menu itself is the product now, not any single item on it.

The Real Arbitrage Is LTV Per Call, Not Price Per Token

Let me put on the pricing hat, because this is where teams get it wrong. They stare at the token rate. They should stare at the value each call produces.

The fundamental question in any business is LTV to CAC. Here it becomes value-per-call versus cost-per-call. A Luna call that resolves a customer ticket is worth the same to your revenue as a Sol call that does the same job. So paying Sol prices for a task Luna handles is pure margin you set on fire.

Now the math. Finout, the FinOps platform, reports the tiers at Sol $5/$30, Terra $2.50/$15, and Luna $1/$6. That is a 40% to 50% cut on a real bill.

But here is the damaging admission. Those savings are not free. To capture them you need routing logic, evaluation, caching, and fallback orchestration. OpenAI says cache reads can save up to 90% and batch jobs get up to 50% off. Those knobs are real money, and they only work if your engineering can actually operate them.

This is the classic golden goose problem. The token discount is a golden egg. The routing system that decides which tier gets which call, that is the goose. Build the goose and the eggs keep coming. Chase eggs one at a time and you stay busy and broke.

There is also a woman-in-the-red-dress here, the shiny distraction. Everyone wants to argue about Sol's Terminal-Bench score. OpenAI reports Sol at 88.8% and Sol Ultra at 91.9%, edging Claude Mythos 5 at 88.0%. Nice numbers. But if your traffic is mostly easy, the benchmark on your hardest tier barely touches your bill.

One more thing pricing people track: the second price. GPT-5.6 launched June 26, 2026 as a gated preview limited to a small group of trusted partners at the U.S. government's request. So there is the published token rate, and there is the practical cost of getting dependable access. It is unclear whether that access gate loosens fast enough for smaller teams. Broad availability was signaled for the coming weeks after the July 9 public launch.

Three signals inside the same shift

2031

Pull back five years. The single-model era is ending, and I do not think it is coming back.

OpenAI named the tiers Sol, Terra, and Luna on purpose. The number 5.6 marks the generation. The names mark durable capability rungs meant to persist and improve across generations. That is a good-better-best ladder built to outlive any one release.

That naming choice is a strategic tell. When your vendor gives you stable rungs, you can build stable procurement around them. You plan a flagship-mid-volume stack once and let each rung improve underneath you. This is counterpositioning against build-it-yourself, because a hybrid rented stack now maps directly to workload criticality.

The asymmetric bet is architecture, not vendor. Whoever owns clean, provider-neutral routing wins the compounding advantage. Gartner forecasts worldwide AI spending to hit $2.59 trillion in 2026, and the FinOps Foundation says 98% of survey respondents now manage AI spend. When that much money flows through inference, a routing layer that shaves 40% is a flywheel, not a feature.

The contrarian view deserves respect. Cloud vendors have segmented cheap and premium compute for years, so maybe this is just clearer packaging of an old pattern. That is fair. The difference is that OpenAI made the tradeoff operational, with named rungs and posted per-token prices you can code against today.

What to Build This Weekend

Do not rebuild your whole stack. Build one tiny router and get your reps in.

First, pull one week of your AI logs. Sort requests into three buckets by difficulty: easy extraction, standard tasks, and hard reasoning. This is your triage map, and it takes an afternoon, no CS degree required.

Second, prototype the routing prompt in Google AI Studio, the free Gemini playground where you can test logic fast. Write a tiny classifier that reads a request and outputs one label: luna, terra, or sol. That label is your router. Test it aggressively, because it will misfire at first, and that is normal.

Third, wire the actual routing in Genspark or a similar agent engine so each label calls the matching tier. Send easy work to Luna at $1 input, standard work to Terra, and only escalate to Sol. Turn on cache reads for repeated context to grab that 90% discount.

Fourth, measure. Track cost-per-resolved-task before and after, not cost-per-token. If your resolution quality holds and your bill drops, the goose is working.

Start with one workflow, the highest-volume one you have. Ship it this week, watch it break, fix it, and let the compounding start.