Two facts from OpenAI’s September 22, 2026 release sit awkwardly next to each other. GPT-6 Luna costs a twentieth of what GPT-6 Sol costs per token. GPT-6 Luna also has the fresher knowledge cutoff: May 18, 2026, against Sol’s April 20, 2026. The cheap model knows about a month more of the world than the expensive one.
That inversion is the most interesting thing in this launch, and almost nobody will talk about it, because the headline is the price cut. Both models land at 50% below their GPT-5.6 predecessors. Sol is $2 per million input tokens and $10 per million output. Luna is $0.10 and $0.50. They replace the GPT-5.6 lineup outright and sit beneath GPT-6 Astra, the flagship that shipped earlier in the same month.
The 20x ratio is the real design decision
Look at the pricing structure rather than the absolute numbers. The input-to-output multiplier is 5x on both models: $2 to $10, $0.10 to $0.50. And the Sol-to-Luna gap is exactly 20x on both input and output.
Clean ratios like that rarely come from a cost model. They come from a product decision about where you want developers to put which workload. A 5x output premium pushes you toward models that read a lot and write a little, which is precisely the shape of most agent work: large context in, small structured decision out. A flat 20x tier gap means the choice between Sol and Luna is never a partial one. You do not save 30% by downgrading a step. You save 95%.
For anyone building multi-step agents, that changes the arithmetic of the control loop. Take an illustrative agent that runs 100 steps, each carrying 50K tokens of accumulated context. That is 5M input tokens: about $10 on Sol, about $0.50 on Luna. The interesting question stops being “which model is best” and becomes “how many of my 100 steps actually need the better model.” Tier gaps this wide make routing the primary architectural concern rather than an optimization you get to later.
A context window that is one shared budget
Both models carry a 1.05M-token context window, split as 922K maximum input and 128K maximum output. Add those and you get 1.05M exactly. That is not a coincidence, and it is a meaningful detail: the input and output limits are not two independent ceilings. They are partitions of one budget.
Practically, this means an agent that hoards context is spending its own headroom. If you fill 900K tokens with tool results and prior turns, you are close to the input ceiling, and the 128K output allowance is the remainder rather than a bonus. Long-horizon agents that accumulate state without pruning will hit the wall from the input side, not the generation side.
The identical window across both tiers is the part that helps architects most. Sol and Luna accept the same context shape, so a routing layer can hand the same assembled prompt to either one without reformatting or truncating. Swapping tiers becomes a parameter change rather than a rewrite. Whether the two models behave equivalently on that context is a separate question, and one that only your own evaluations will answer.
What the mismatched cutoffs suggest
Back to the inversion. If Luna were a distillation of Sol, or if both were pruned from one parent checkpoint, you would expect them to inherit the same data snapshot. They did not. Sol stops at April 20, Luna at May 18.
I want to be careful here, because OpenAI has not published the training details, and what follows is inference rather than fact. Different cutoffs imply the two models were trained on separately assembled corpora, likely on staggered schedules. Whatever relationship exists between them, it is not a single lineage where one model is mechanically derived from the other at the same point in time.
The operational consequence is concrete regardless of the cause. If you route a task to Sol for its capability tier and it reasons about something from early May 2026, it is working from a gap that Luna does not have. Recency and capability are usually correlated in a model family. In this pair, they run in opposite directions. Any agent that mixes tiers within a single task now has two different world-knowledge boundaries in play, and no obvious signal at runtime telling it which one it just used.
What I would measure first
Three things, before committing to either model in production:
- Output-token behavior under load. A 5x output premium makes verbosity expensive, so measure actual tokens generated per step, not just quality of the final answer.
- Where Luna fails and Sol succeeds on your own tasks. With a 20x gap, the correct default is Luna everywhere until you have evidence of a specific failure mode.
- Knowledge-boundary sensitivity. If your workload touches anything from spring 2026, the cutoff difference is a correctness issue rather than trivia.
The price cut will get the attention. The structural details of this release, one shared context budget, an exact 20x tier gap, and two models whose knowledge recency runs against their cost, are what will actually shape how agents get built on top of them.
🕒 Published: