Remember when GPT‑6 Astra landed earlier this month and half the bot-building community spent a weekend swapping model strings, re-tuning prompts, and then quietly discovering their token bill had opinions about it? That cycle barely finished. On September 22, 2026, OpenAI dropped two more models into the GPT‑6 lineup — GPT‑6 Sol and GPT‑6 Luna — as direct replacements for the GPT‑5.6 versions, at half the price.
So here I am, back in the routing layer again. But this time the change is worth the churn, and not for the reason the headlines are chasing.
The numbers that actually matter
Sol runs $2 per million input tokens and $10 per million output tokens, down 50% from GPT‑5.6 Sol’s promotional pricing. Luna runs $0.10 per million input and $0.50 per million output. Both are described as the cheaper, faster siblings to Astra, the flagship.
Look at the ratio rather than the absolute figures. Luna is twenty times cheaper than Sol on both input and output. That is not a discount, that is a different tier of product. Run the math on a moderately busy bot pushing 30 million input tokens and 3 million output tokens a month:
- Sol: $60 input + $30 output = $90
- Luna: $3 input + $1.50 output = $4.50
Same workload. Two orders of magnitude difference in how much you care about it. If you have been rate-limiting your users, trimming conversation history, or skipping retrieval passes to protect your margins, Luna changes which of those constraints are real and which were just cost anxiety wearing an engineering costume.
A context window with a split personality
Both models carry a 1.05M-token context window. The part builders should read twice: 922K of that can be input, with 128K as the max output. Those two numbers add up to the total, which tells you the window is not one undifferentiated pool you can spend however you like.
Practically, that means two separate things to design around. On the input side, 922K is enough that “stuff the whole document set in there” becomes a legitimate architecture rather than a joke you make in a design review. On the output side, 128K is a hard ceiling. Any long-generation job — full codebase scaffolds, bulk translation, report generation across hundreds of sections — still needs chunking and stitching logic. The window got enormous; your output loop did not get to retire.
I would also resist the urge to route everything through max context just because it exists. Input tokens are still billed per million. A 900K-token prompt on Sol costs you roughly $1.80 before the model writes a single word. On Luna, nine cents. The pricing gap and the context ceiling interact, and the interesting designs live at that intersection.
The knowledge cutoff quirk worth writing down
This is the detail I nearly missed. The cutoffs differ between the two models: April 20, 2026 for Sol, May 18, 2026 for Luna. The cheaper model has the more recent knowledge.
That breaks an assumption a lot of us carry without examining it — that price tier correlates with recency, capability, and freshness all at once. It does not. If your bot answers questions where a month of world knowledge matters, the cheap model is the one with better coverage. If you built a fallback chain that degrades from expensive to cheap on failure, your fallback may be returning fresher answers than your primary. Worth knowing before a user points it out for you.
How I am rewiring things
My plan, in rough order:
- Make Luna the default, not the fallback. Route to Sol on escalation instead of demoting to Luna on failure. Most bot traffic is classification, extraction, summarization, and routing — not the hard stuff.
- Measure before assuming. Run your existing eval set against both. A twenty-times price gap justifies a real afternoon of testing, and any quality claim you make without numbers is just vibes.
- Pin model versions explicitly. These are direct replacements for the GPT‑5.6 lineup, which is exactly the kind of substitution that silently changes behavior in production.
- Log cutoff-sensitive queries separately. With the two cutoffs a month apart, you want visibility into which model answered anything time-dependent.
The part nobody puts in a launch post
Cheaper models do not simplify your architecture. They move the difficulty. When inference costs pennies, the hard problems become routing accuracy, eval coverage, latency budgets, and knowing which model handled which request when something goes wrong. That work does not get cheaper by 50%.
Still, I will take it. A month ago I was designing around cost. This week I get to design around the problem. That is a better place to be building bots from, and it is the first thing I would tell anyone still sitting on a GPT‑5.6 model string.
🕒 Published: