GPT-5.6 Luna Pricing: $1.58 per 1,000 Chatbot Replies

OpenAI cut GPT-5.6 Luna’s API price by 80% on 30 July 2026, taking it from $1.00 to $0.20 per million input tokens and from $6.00 to $1.20 per million output tokens. Most of the coverage framed that as a story about the AI price war. For anyone running a chatbot on a WordPress site, it is a much more concrete story than that, and it has a number attached: a thousand chatbot replies on Luna now cost about $1.58.

That figure is not a vendor estimate. It comes from measuring what one reply on this site actually sends — 7,000 input tokens and roughly 150 output tokens — and multiplying it by OpenAI’s published rate card. The interesting part is not that Luna got cheaper. It is that the gap between the cheapest and the most expensive model inside the same family is now roughly twentyfold, which is a far bigger lever than choosing between vendors.

What actually changed, and when

Two separate events repriced the GPT-5.6 family inside four weeks, and they are frequently confused with each other:

• 30 July 2026 — Luna dropped 80% ($1.00/$6.00 to $0.20/$1.20) and Terra dropped 20% ($2.50/$15.00 to $2.00/$12.00). Sol, the flagship, was explicitly not included. OpenAI attributed the cuts to serving efficiency gains rather than competitive pressure.
• 21 August 2026 — Sol was cut separately, from $5.00/$30.00 to $4.00/$20.00, as promotional pricing guaranteed “at least through November 21, 2026”. We covered that one in detail in GPT-5.6 Sol pricing and the November 21 catch.

The distinction matters for budgeting. The Luna and Terra cuts carry no expiry date attached to them. The Sol cut does. If you build a cost model on Sol’s current $4.00 input rate, you are building on a promotional rate with a stated review date; if you build it on Luna’s $0.20, you are not.

The full GPT-5.6 rate card

Read directly from OpenAI’s pricing page on 30 August 2026, per million tokens:

One detail worth noticing before the arithmetic: Luna and Terra share an identical 1:6 ratio between input and output pricing. Sol’s is 1:5. That sounds like trivia, but it means the relative economics of a verbose bot versus a terse one are the same on Luna and Terra, and only slightly different on Sol.

What a reply actually costs

A retrieval chatbot’s cost per reply is dominated by what you send it, not by what it says back. On this site, one reply assembles six retrieved content chunks at the plugin’s 4,000-character default, plus a 3,478-character system prompt. That is 27,478 characters, or roughly 6,870 tokens — call it 7,000. A typical answer runs about 150 tokens.

Those two numbers give a 47:1 input-to-output ratio, and that ratio is what makes the following table look the way it does.

Two things fall out of this that are worth more than the headline percentage:

The tier you pick matters about twenty times more than the cut you read about. Moving from Sol to Luna saves $29.42 per thousand replies. The 80% cut itself saved $6.32 per thousand. Both are real, but the first is nearly five times larger than the second, and it is available to you today rather than on OpenAI’s schedule.

The absolute numbers are small enough to change how you think about the feature. A WordPress site fielding 1,000 chatbot conversations a month is looking at under two dollars of model cost on Luna. At that level the API bill is no longer the thing worth optimising — your hosting probably costs more. This is a genuinely different situation from the one we described in our original write-up on chatbot running costs, and it is worth revisiting your assumptions if you priced this out a year ago.

Why output price cuts keep disappointing people

When Sol was cut in August, the headline number was 33% — that was the output-side reduction, from $30.00 to $20.00. The input side fell 20%. Because a retrieval chatbot sends 47 tokens for every one it returns, the blended saving came out at 21.5%, much closer to the input figure than the headline.

The chart makes the general case. Input tokens are 88.6% of the bill on Luna and Terra, and 90.3% on Sol. Any pricing announcement that leads with an output-side reduction is describing a change to roughly a tenth of your invoice. When you read the next one, find the input number before you form a view — it is the one that will show up on your card. We made the same argument from the other direction when comparing per-resolution pricing to raw tokens in our analysis of Intercom Fin’s pricing.

The caching lever, and its honest limit

Since input dominates, cached input looks like the obvious next lever. Luna bills cached input at $0.02 against $0.20 — a tenfold reduction on the expensive side of the ledger. In practice the saving is much smaller than that ratio suggests, and it is worth being precise about why.

Prompt caching only pays off on the portion of your input that is stable between requests. For a retrieval chatbot, that is the system prompt — about 870 tokens here. The remaining ~6,130 tokens are retrieved content that changes with every question, which is the entire point of retrieval.

So caching is worth about 10% on a retrieval bot and up to 80% on a bot with a fixed prompt and no retrieval. Both are true; which one applies to you depends entirely on your architecture. If someone quotes you the 80% figure for a RAG chatbot, they have not thought it through.

Is the cheap tier actually good enough?

This is where an honest article has to stop reciting prices. We have not benchmarked answer quality across the three tiers, and we are not going to claim a result we did not measure. OpenAI positions Luna as its most cost-efficient model rather than its most capable, and that framing should be taken at face value.

What can be said usefully is which parts of the job are sensitive to model strength. Retrieval-grounded support answers — where the correct response is largely contained in the chunks you supplied — lean on the model for summarising and tone rather than for reasoning, and that is the workload cheap tiers handle best. Multi-step reasoning, ambiguous product comparisons, and anything where the model has to decide what the user really meant are where a stronger tier earns its cost.

The practical approach is to run the tiers side by side on twenty of your own real questions before committing. A twentyfold price difference justifies an afternoon of testing, and your questions are the only benchmark that matters. If you are running MxChat, the model is a dropdown in the settings, so running the same twenty questions across all three tiers costs you under a dollar in total, roughly two-thirds of which is the Sol run.

What the search results say — and the one figure that is wrong

Because pricing pages go stale quickly, we ran a census of every organic result Google returns for gpt-5.6 luna pricing and checked what rate each one states.

The result is reassuring, and worth reporting as such rather than dramatised: four of the six organic results state the current rate correctly. Google returned six organic results in total for this query, at ranks 2, 3, 4, 6, 7 and 11 — nothing organic came back at rank 1, 5, 8, 9 or 10.

The one wrong figure is a Reddit comment claiming “$0.1 input and $0.6 output” — exactly half the real price, and clearly a user misremembering rather than a source anyone should be pricing against. The sixth result is OpenAI’s own GPT-5.6 launch post from 9 July, which predates the cut and quotes no per-token rate in its snippet at all. That is not an error so much as an artifact of a dated announcement outranking a live rate card.

The staleness that is worth flagging sits off the SERP entirely. At least one dedicated API pricing tracker, stamped as updated on 29 August 2026, currently lists all three GPT-5.6 models at their pre-cut rates — Luna at $1.00/$6.00, Terra at $2.50/$15.00, Sol at $5.00/$30.00. A tracker updated yesterday can still be a month behind on every row, and being 5x wrong on Luna is the difference between a $1.58 bill and a $7.90 one.

The rule that follows is boring and it works: price from the vendor’s own rate card, and check the date on anything else. The same discipline applies to model availability, not just price — retirements catch people the same way, which is what we covered in our piece on OpenAI model deprecations.

What to do with this

1. Find out what your bot actually sends. Your chunk count and chunk size drive nearly 90% of your bill. Six chunks at 4,000 characters is a default, not a law — if four chunks answer your questions as well, you have cut a third of your cost without changing model.
2. Default to the cheap tier and escalate only if you can show it failing. The 20x spread inside one family means tier choice dominates almost every other cost decision you will make.
3. Check whether a rate you rely on has an expiry. Luna’s and Terra’s do not. Sol’s is guaranteed only through 21 November 2026.
4. Cache the system prompt, but budget for 10% rather than 80% unless your prompt is genuinely static.
5. Re-read the rate card quarterly. Three of the GPT-5.6 family’s rates changed within four weeks this summer.

Frequently asked questions

How much does a WordPress chatbot cost to run on GPT-5.6 Luna?

About $1.58 per 1,000 replies, assuming a retrieval setup that sends roughly 7,000 input tokens and returns 150 output tokens per reply. Sites with smaller prompts or fewer retrieved chunks will pay less; the input side is where nearly all of it lives.

Is the GPT-5.6 Luna price cut permanent?

OpenAI announced the 30 July reduction without an expiry date, unlike the Sol cut of 21 August, which is described as promotional pricing available “at least through November 21, 2026”. Neither guarantee binds forever, but only one of them has a date on it.

Why is Luna so much cheaper than Sol for the same conversation?

Luna’s input rate is $0.20 per million tokens against Sol’s $4.00, a twentyfold difference, and input is roughly 90% of a retrieval chatbot’s bill. The output gap contributes very little by comparison.

Does prompt caching cut my bill by 90%?

Only if nearly all of your input is identical between requests. Cached input is billed at 10% of the standard rate, but a retrieval chatbot changes most of its input every turn, so the realistic saving is closer to 10% overall.

Which model should I use with MxChat?

Start on the cheapest tier that answers your own questions acceptably and move up only when you can point at a failure it caused. MxChat lets you switch models from the settings screen, so the comparison costs almost nothing to run — see the documentation for the model options, or MxChat Pro if you want the full provider list.

All prices in this article were read from OpenAI’s published API pricing page on 30 August 2026 and cross-checked against contemporaneous reporting of the 30 July announcement. Rates change; check the date on this page against the date on the rate card before relying on it.