GPT-5.6 Sol Pricing: $4/$20 and the Nov 21 Catch

On 21 August 2026, OpenAI cut the API price of GPT-5.6 Sol. Input went from $5.00 to $4.00 per million tokens, output from $30.00 to $20.00. That is a real cut and, if you run a WordPress chatbot on Sol, it lands on your next invoice whether or not you noticed it happen.

What I did not expect, when I went to check the number, was that page one of Google cannot tell you what GPT-5.6 Sol costs. Three of the nine organic results still quote the pre-cut price. Three are about a different price event entirely. Exactly one of them carries the qualifier that OpenAI’s own documentation puts on the number, and it is not the one you would guess.

So this is two things: what the new rates actually are, verified at OpenAI’s own rate card rather than from the coverage, and a short census of how badly the web has that wrong right now. Then the part that matters for a WordPress site — what a chatbot reply actually costs at the new numbers, using this site’s own live configuration rather than a made-up example.

What actually changed on 21 August

Two figures moved, both on the standard short-context tier:

The input cut is 20%. The output cut is 33%. Coverage generally led with one or the other, which is where the “OpenAI cuts prices by more than 20%” and “OpenAI cuts prices by up to 33%” headlines both come from — they are the same event, described from opposite ends.

I confirmed the change two ways rather than trusting a summary. OpenAI’s developer pricing page lists gpt-5.6-sol at $4.00 input and $20.00 output today. Amazon’s Bedrock “what’s new” announcement, dated 21 August 2026, gives the same pair and states the reduction explicitly as 20% on input and 33.3% on output — an independent origin, from a party that has to bill correctly against it.

Here is the full GPT-5.6 family at the rates showing today, short-context tier:

Terra and Luna did not move on 21 August. Their current rates come from an earlier cut, at the end of July. Sol is the one that changed last week.

“At least through November 21” is not a deadline

This is the detail almost every write-up drops, and it inverts the advice.

The sentence on OpenAI’s pricing page reads: “GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026.” Amazon’s announcement uses effectively the same construction — “This promotional pricing is available at least through November 21, 2026.”

“At least through” is a floor, not an expiry. It is a commitment that the price will not go up before 21 November. It says nothing whatsoever about what happens on 22 November. The rate could persist, could become permanent, or could revert.

Most of the coverage compressed this to “through November 21,” which reads as a countdown — as though there is a date on which your bill jumps back to $5/$30. That framing produces exactly the wrong behaviour: rushing to consume cheap tokens before a deadline that has not been announced, or building a migration plan for a reversion nobody has committed to.

The correct reading is duller and more useful. You have a guaranteed floor price until 21 November. After that, unknown. Budget both numbers and get on with your day.

Page one does not agree with itself

Before writing any of the above I ran the live SERP for gpt-5.6 sol pricing to see what a site owner researching this would actually find. The result is worth publishing on its own.

Nine organic results came back, at absolute ranks 2 through 11. Classified on the price each one displays in its own snippet:

• Three show the current price. OpenAI’s own docs at #2, an AI pricing tracker at #8, and Reuters at #10.
• Three still show $5 / $30. One of them is dated 20 August — published the day before the cut, and now simply wrong. Another also lists Terra at $2.50/$15, which does not match the current card either.
• Three are about something else. Two concern an OpenRouter-specific promotional rate of $2/$10, which is a different provider’s pricing for the same model. The third is OpenAI’s own July announcement of an earlier Sol price drop — a real page about a real event, seven weeks stale relative to the query.

Two caveats on that census, because they matter. First, I classified the snippet Google displays, not the current body of each page — a publisher may have updated the page while Google still serves an older snippet. For a searcher, the snippet is what they see, which is why it is the thing worth measuring, but it is not a claim about those publishers’ current content. Second, no organic result was returned at rank 1 or rank 4; I am reporting that as measured, not explaining it.

The one result carrying OpenAI’s actual qualifier is Reuters, whose snippet specifies “for standard short-context use.” Which brings us to the part nobody covered at all.

The rate card has eight prices, not two

“Sol is $4/$20” is a compression of a table with eight cells in it. The GPT-5.6 models are billed at two different tiers depending on how much you put in the request:

Above 272,000 input tokens, the request bills at the long-context tier: 2× on input, cached input and cache writes, and 1.5× on output. The multipliers apply to the whole request, not just the tokens above the line. Those multipliers reconcile exactly against OpenAI’s published figures — $4 becomes $8, $0.40 becomes $0.80, $20 becomes $30.

There is a tempting story here, and I want to be clear that I checked it and it is not true. The tempting story is: “the price cut is a mirage, because long-context output is $30, exactly what you used to pay.” Arithmetically that is correct. Practically it is irrelevant to a WordPress chatbot, and saying otherwise would be manufacturing alarm.

Here is the scale check, measured on this site’s own live configuration rather than assumed. MxChat retrieves 6 content chunks per reply. The plugin’s default chunk size is 4,000 characters, roughly 1,000 tokens. The system prompt on this install is 3,478 characters, about 870 tokens. Add the visitor’s question and a little conversation history and a typical request is somewhere near 7,000 input tokens.

That is 2.6% of the 272,000-token threshold. A WordPress chatbot answering from site content sits roughly forty times below the long-context tier. It pays $4/$20, full stop. The long-context rates matter if you are feeding entire codebases or book-length documents into a single call; they do not matter here.

What a reply actually costs now

Using those measured numbers — about 7,000 input tokens per reply, and an output length governed by a system prompt that instructs the bot to answer in one to three short sentences, call it 150 tokens:

The cut is worth about 21% off a Sol-powered chatbot’s token bill — not 33%, and not 20% either.

The reason is worth understanding, because it generalises to every pricing announcement you will read this year. A retrieval chatbot is enormously input-heavy. Roughly 7,000 tokens go in; roughly 150 come out. That is a 47:1 ratio. So the headline number — the 33% output cut, the bigger and more quotable of the two — moves almost nothing on your bill. The unglamorous 20% input cut is doing nearly all the work.

If you take one operational habit from this piece, make it that one: when a vendor announces a price change, weight the input and output cuts by your own actual token ratio before deciding whether you care. For a summarisation or code-generation workload the conclusion reverses entirely. I went through the same arithmetic from the other direction when I measured what Intercom Fin’s per-resolution pricing works out to in raw tokens, and the ratio was the whole story there too.

One more lever sitting in that table: cached input is $0.40 against $4.00, a tenth of the price. Your system prompt is identical on every single request, and on a site with a stable knowledge base many retrieved chunks repeat across conversations. That is the largest cost reduction available to most WordPress chatbot operators, and it is bigger than the price cut this article is about.

What to actually do about it

1. Check which model your plugin is actually calling. Not which one you chose last spring — which one is configured today. MxChat ships support for all three GPT-5.6 tiers, and the gap between Sol and Luna in the table above is a factor of 19.
2. Budget both rates past 21 November. Not because a reversion is coming, but because no one has said it isn’t. If your annual forecast assumes $4/$20 in perpetuity, note the assumption where the next person will find it.
3. Do not migrate models over a promotional rate. Three months of guaranteed pricing is not a reason to re-tune prompts and re-test outputs against a different model. Migration costs are real; the delta here is single-digit dollars for most sites.
4. Turn on caching before you shop for a cheaper model. A tenth-price input path on a repeated system prompt beats a 20% headline cut, and it does not change a single answer your visitors see.
5. Re-check the number at the origin. Given that a third of page one is quoting a price that expired eight days ago, the only reliable source is the vendor’s own rate card.

Frequently asked questions

Does the GPT-5.6 Sol price go back up on 21 November 2026?

Unknown, and OpenAI has not said it will. The published wording is that promotional pricing is available “at least through” that date, which guarantees a floor rather than announcing an expiry. Plan for both outcomes; don’t act as though a reversion has been scheduled.

Does the price cut apply if my chatbot uses a long context?

Almost certainly yes, because your chatbot is not using a long context in OpenAI’s sense. The long-context tier begins above 272,000 input tokens per request. A WordPress chatbot retrieving a handful of content chunks sends something on the order of 7,000. You are paying the short-context rate.

Should I switch from Sol to Terra or Luna to save money?

That is a quality decision with a cost consequence, not a cost decision. The right way to run it is to take twenty real questions your visitors have actually asked, run them against each tier, and read the answers side by side. If Luna answers your visitors’ questions correctly, the table above says it costs about 5% of Sol. If it doesn’t, no price makes it the right choice.

Where do I see what my chatbot is actually spending?

Your model provider’s usage dashboard is the authoritative number, because it counts tokens you may not be thinking about — retries, system prompt, retrieved context. Multiply your monthly reply count by the per-reply figures above for a sanity check; if the two disagree by a lot, the usual culprit is retrieval returning more or larger chunks than you expected. MxChat’s documentation covers where chunk size and source count are configured.

Is this the same as the OpenRouter price cut people were discussing?

No, and conflating them is why two of the nine results on page one are misleading. OpenRouter listed a promotional $2/$10 rate for Sol in mid-August. That is a reseller’s pricing on their own platform, not OpenAI’s API rate. If you call OpenAI directly — which is what a WordPress plugin with your own API key does — the rates in this article are the ones that apply to you.

The short version

GPT-5.6 Sol is $4.00 in and $20.00 out per million tokens, down from $5.00 and $30.00 on 21 August 2026, guaranteed at least through 21 November. For a retrieval chatbot the real saving is about 21%, driven almost entirely by the input cut rather than the larger output cut, because a chatbot reply is roughly 47 times more input than output. The long-context tier exists but starts nearly forty times further out than any WordPress chatbot will ever reach.

And a third of Google’s first page will tell you the price is still $5/$30. It isn’t.

If you are sizing this alongside the other providers, I have run the same exercise on Claude Sonnet 5, whose scheduled September increase was cancelled outright, and on Gemini 3.7 Flash’s half-price window — which, unlike this one, does carry a stated end date. Sol is also the most expensive way to run this workload — the same measured reply costs 19.6x less on GPT-5.6 Luna, which is the comparison that actually moves a WordPress chatbot’s bill. For the broader picture of what running an AI chatbot on WordPress costs across plugins and providers, there’s a full pricing guide here. And if you want to switch models without touching code, that is what MxChat Pro is for — the model is a dropdown, so the arithmetic above is a setting, not a migration.