6 October 2026 · AI & LLMs · 5 min read
Here's the detail that made me stop and check the numbers twice: GPT-6.1 Sol costs $2 per million input tokens and $10 per million output. GPT-5.6 Sol — the model it's named after, the one OpenAI is still running a "discount" on through at least November 21 — currently costs $4/$20, even with that discount applied. The new model is half the price of the old model's clearance rate. Nobody launches a next-gen tier priced under last gen's fire sale by accident.
GPT-6 Astra took the spotlight when it launched September 3 at $10/$50 per million tokens. Quietly filled in underneath it: GPT-6.1 Sol and GPT-6 Luna, both confirmed on OpenAI's own pricing page this week. Standard (short-context) rates, per million tokens:
| Model | Input | Output | vs GPT-5.6 equivalent |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 | 2.5× GPT-5.6 Sol's discount rate |
| GPT-6.1 Sol | $2.00 | $10.00 | 50% cheaper than GPT-5.6 Sol's own $4/$20 promo |
| GPT-6 Luna | $0.10 | $0.50 | same as GPT-5.6 Luna's July repricing |
So the family isn't a clean "everything costs more" refresh. Astra is the expensive new flagship. Luna held flat. Sol — oddly — got cheaper on arrival than the model it replaces, discount included. If you're still budgeting against GPT-5.6 Sol's sticker price, you're overpaying for no reason; if you're budgeting against its temporary discount, you're about to be pleasantly surprised when that discount expires and GPT-6.1 Sol is still there at half the rate.
All three new tiers share the same structural quirk, and it's the kind of thing that doesn't show up until a real workload hits it: cross 272,000 input tokens on a single request and the entire request — not just the overage — gets billed at a higher rate. Input roughly doubles, output goes up 50%:
| Model | ≤272K tokens (in/out) | >272K tokens (in/out) |
|---|---|---|
| GPT-6 Astra | $10.00 / $50.00 | $20.00 / $75.00 |
| GPT-6.1 Sol | $2.00 / $10.00 | $4.00 / $15.00 |
| GPT-6 Luna | $0.10 / $0.50 | $0.20 / $0.75 |
I wrote about the same mechanic on Grok 4.7's 200K-token cliff last month — this is the second major vendor I've found doing whole-request repricing instead of per-token overage. It's becoming a pattern worth architecting around, not a one-off.
Say you're running a document-analysis pipeline: 500,000 input tokens per document (well into long-context territory), 5,000 tokens of summary output. Per request:
| Model | Cost per request | 2,000 requests/mo |
|---|---|---|
| GPT-6 Astra | $10.375 | $20,750 |
| GPT-6.1 Sol | $2.075 | $4,150 |
| GPT-6 Luna | $0.104 | $208 |
Same document, same pipeline, same month: $20,750 or $208 depending which tier you default to — a 100x spread for output quality that doesn't differ 100x. For a short chat-style exchange (2,000 input / 500 output tokens, nowhere near the cliff) the gap is just as real in relative terms: Astra runs $0.045 a turn, GPT-6.1 Sol $0.009, Luna under half a cent. At chatbot volume — say 200,000 turns a month — that's $9,000 on Astra versus $1,800 on Sol versus $90 on Luna.
Default to Luna. Seriously — $0.10/$0.50 is cheap enough that most classification, extraction and routing tasks should never see Astra's rate card. Reserve Astra for the subset of requests that actually need its reasoning depth, and route everything else down. If your pipeline ever touches documents near 272K tokens, don't let it happen by accident: chunk or summarize before the line on purpose, the same advice I gave for Grok's 200K cliff, because crossing it silently doubles your input cost on the whole request, not just the tail end. And if you're still pricing against GPT-5.6 Sol because that's what's in your code — check again. Its $4/$20 rate is explicitly temporary (through at least Nov 21), and GPT-6.1 Sol already beats it on a rate that isn't going anywhere.
New to usage-based LLM pricing? Start with the free API-cost guides. For the other vendor doing whole-request cliff pricing, see Grok 4.7's 200K-token cliff.
Rates verified against OpenAI's official API pricing page and cross-checked against independent coverage, 6 October 2026. Reference estimates — confirm current pricing and your actual token mix before budgeting.