On June 26, OpenAI opened a limited preview of GPT-5.6, and the new flagship model — Sol — promptly posted a 91.9% on Terminal-Bench 2.1 in its “Ultra” reasoning mode. That edges Claude Mythos 5 and GPT-5.5, both at 88.0%. The Hacker News threads are, predictably, a mix of awe and dread. Will this replace junior devs? Is the junior dev already dead?
Buried beneath the benchmark scoreboard is a number that got less attention: $30 per million output tokens. That’s the price for Sol Ultra. Input tokens are $5 per million. The “balanced” Terra tier runs $2.50/$15. Luna, the cheap one, is $1/$6.
Put another way, the best software-writing machine humanity has ever built costs roughly thirty times more to query than its own budget-tier sibling. And that’s the column.
Not the benchmarks. Not the “will it replace us” parlor game. The pricing. Because the pricing tells you who OpenAI actually expects to use this thing.
The Real Announcement Is a Two-Tier Intelligence Market
$30 per million output tokens is not a consumer price. It’s not a prosumer price. It’s a “send this to the lawyers before you click” enterprise procurement price. A single non-trivial Codex session — the kind where you’re iterating on a pull request across multiple files, refining, re-prompting — could burn through a lunch tab without anyone noticing.
And Sol Ultra is the only model in the lineup that clears 90% on a coding benchmark. So the state of the art is now, practically speaking, gated by a credit card limit.
This isn’t inherently sinister. Compute costs money. Frontier models are expensive to run. The interesting question is what happens downstream. If the best coding AI is a luxury good, who gets to use it? Well-funded startups, sure. Big tech, obviously. But the solo developer bootstrapping a SaaS product? The small manufacturer in Ohio who needs a custom inventory tool? The public school IT admin trying to patch a legacy system? They’re priced into Luna, which is fine — Luna is likely excellent — but it isn’t going to score 91.9% on anything.
We’ve spent two years arguing about whether AI will democratize software development. The pricing suggests the opposite: it’s creating a tiered system where the genuinely powerful tools sit behind a velvet rope.
The “Tier” Branding Is Doing More Work Than Anyone Admits
OpenAI’s naming scheme — Sol, Terra, Luna — is cosmetic, but it’s not meaningless. Cosmetic choices at this scale are never meaningless. The company says the tiers represent “durable capability levels that can advance on their own cadence.” What that means in practice is that Sol will always be the expensive one. It will always be the frontier. Terra and Luna will play catch-up, forever.
This is a departure from the old model, where GPT-4 got cheaper over time and eventually trickled down to the free tier. The new architecture implies permanent stratification. You don’t wait for Sol to get cheap; you wait for an improved Luna that approximates what Sol could do six months ago.
To be fair, that’s how most markets work. First-class seats don’t get cheaper; they add more legroom. The interesting break is that this is a market for intelligence. Not compute, not storage, not bandwidth — raw problem-solving capacity. And we’re pricing it like a premium airline cabin.
One engineer I know, who works on developer tools at a mid-size company and Slacked me during the announcement, put it bluntly: “They’re not selling a model. They’re selling a membership tier for competence.”
He’s not wrong.
The Junior Dev Is Fine — It’s the Middle Tier That Should Worry
The standard narrative is that AI coding tools will replace entry-level engineers. But Sol Ultra is priced for senior-level work. The economics point the other way: if you’re a company, you don’t pay $30 per million tokens to replace a $70,000 junior dev. You pay it to augment a $250,000 senior engineer and make them 30% faster. That’s the arbitrage.
The group that should be nervous is the mid-career developer who’s good but not great — the one who writes solid code but doesn’t architect systems, who can ship features but doesn’t set direction. That’s the labor that starts to look like a rounding error next to a $30/MTok API call that ships in minutes.
And that’s a different conversation than the one we’ve been having. The threat isn’t from below. It’s from above — from the model that does what the middle tier does, but faster, and for a marginal cost that makes headcount look like a legacy expense.
None of This Is a Complaint
It’s easy to read a column about pricing tiers and assume the writer wants the government to step in and cap the price of frontier model tokens. That would be a terrible idea, and it’s not what I’m arguing.
What I’m arguing is that the pricing architecture OpenAI chose this month is a signal. It tells us the company believes the frontier is not going to commoditize quickly. It tells us they think the gap between Sol and Luna is durable enough to monetize indefinitely. And it tells us that the old dream — AI as a rising tide that lifts all boats equally — is being quietly shelved in favor of something more recognizable: a premium product for premium customers, with a perfectly adequate budget option for everyone else.
That’s not a crisis. It’s just the end of a certain kind of naivete. The best coder in the world costs $30 per million tokens. That’s not expensive for what it does. But it’s expensive enough that not everyone gets to use it.