OpenAI added two models to its GPT-6 family on September 22: GPT-6 Sol, aimed at complex coding and agentic work, and GPT-6 Luna, its efficient tier for high-volume tasks. On price, OpenAI’s announcement, as posted in the Announcements category of its Developer Community forum, says the company made “caching and inference more efficient” and is passing the savings on as “50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.”
The 50% figure checks out against the list prices. For anyone re-running a cost estimate, the useful questions are what that 50% is measured against and which line items moved by how much.
The list prices side by side
These are the Standard text-token prices per 1M tokens from OpenAI’s model pages for GPT-6 Sol, GPT-6 Luna, GPT-5.6 Sol and GPT-5.6 Luna:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-5.6 Sol | $4.00 | $0.40 | $20.00 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
Sol is a clean 50% cut on every line. Luna’s input and cached-input prices also halve, but its output price drops from $1.20 to $0.50, which works out to about 58% lower. For Luna jobs that generate a lot of text relative to what they read, the saving is bigger than the headline.
The GPT-6 pages also itemise cache writes: $2.50 per 1M tokens for Sol and $0.125 for Luna, or 1.25x the uncached input rate. Both GPT-5.6 pages state the same 1.25x rule, so this is not a new charge, just a more visible one. The GPT-6 pages also list the pricing modifiers: regional processing adds 10% where available, Batch and Flex are billed at 50% of Standard, Fast mode at 2x, and EU data residency is available only with Standard processing.
Beyond price
According to the same pages:
- Both generations have a 1,050,000-token context window and 128,000 max output tokens.
- On both, prompts over 272K input tokens are billed at 2x input and 1.5x output for the full request. The GPT-6 pages add that the 2x also applies to cache rates. The GPT-5.6 model pages don’t mention cache rates here, but OpenAI’s pricing page, in its table of Daybreak models, lists GPT-5.6 Sol’s long-context rates as $8.00 input, $0.80 cached input, $10.00 cache writes and $30.00 output, so Sol’s cache rates double too. No long-context cache rate for GPT-5.6 Luna appears in these sources.
- The knowledge cutoff moves from February 16, 2026 on both GPT-5.6 models to April 20, 2026 for GPT-6 Sol and May 18, 2026 for GPT-6 Luna.
“Promotional” is doing some work in that sentence
OpenAI measured the cut against GPT-5.6 promotional pricing, and the GPT-5.6 Sol page says that promotional pricing “is available at least through November 21, 2026.” The comparison baseline was already a discounted rate, at least for Sol; the GPT-5.6 Luna page carries no promotion note. The pages don’t say what GPT-5.6 Sol costs after the promotion, so the gap after the promotion ends can’t be calculated from these sources.
The GPT-6 model pages carry no promotion end date. VentureBeat reports that an OpenAI spokesperson confirmed the new Sol and Luna rates are permanent prices, not promotional or introductory pricing. That is a reported statement rather than documentation, but it points the same way.
Caching: the same discount ratio
Cached input is 10% of the uncached input rate on all four models. Because input and cached input both fell by 50%, a cache hit saves the same share of the input bill on GPT-6 as it did on GPT-5.6. The rules are in OpenAI’s prompt caching guide: the entire rendered prefix has to match, and for GPT-5.6 and later the minimum cacheable prefix is 1,024 visible input tokens and a cached prefix stays eligible for at least 30 minutes after its last write or reuse. On GPT-6 models, the guide says reasoning effort can be changed between responses with a configuration_update input item while the top-level reasoning.effort stays at its original value, because changing that top-level setting “can rewrite instructions in the hidden system instructions.”
What to take from it
My reading, not OpenAI’s:
- If GPT-5.6 Sol or Luna was chosen on cost, list price alone now favours GPT-6. At Standard rates, the saving is exactly 50% for Sol at any prompt length. For Luna it is 50% or more below 272K input tokens, rising with the share of output; above 272K, the cached-token comparison for Luna can’t be made from these sources.
- Whether quality holds for a given task is a separate question that the price tables don’t answer.
- The newer knowledge cutoffs are a small but real difference for anything that leans on recent facts.
What the sources don’t say: what GPT-5.6 Sol costs after its promotion ends, when or whether GPT-5.6 Sol and Luna will be retired, and how GPT-6 Sol and Luna compare with other vendors’ models on independent evaluations.