Pricing and documentation verified: 23 August 2026. Compiled from OpenAI’s official developer documentation — the GPT-5.6 Sol model page, the API pricing page and the API changelog — all read on 23 August 2026. We do not hold an OpenAI API account, have not run any of these models, and have not seen an invoice. Every rate and every rule below is a published vendor number or a verbatim vendor sentence. Every percentage, ratio, break-even and worked example built on top of them is our own arithmetic and is labelled where it appears.

Compiled and fact-checked by the AI Tools Worth editorial team. Corrections: contact page.

What actually changed on 21 August

OpenAI cut the price of its flagship model. The API changelog entry dated 21 August 2026 and the model page carry the same sentence, published verbatim:

“GPT-5.6 Sol costs $4 per million input tokens and $20 per million output tokens, a 20% reduction in input pricing and a 33% reduction in output pricing.”

And, separately on the same pages:

“GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026.”

Those two sentences are the entire published basis for everything that follows. Note the hedge in the second one: “at least through”, not “until”. OpenAI has committed to a floor on the duration and to nothing else. It has not published what the price becomes on 22 November, or whether it changes at all.

The stated reductions pin the previous rates arithmetically: $4 is 20% below $5, and $20 is 33.3% below $30. That matches the rates GPT-5.6 Sol launched with, and matches wire coverage of the announcement. Third-party pricing summaries published before 21 August still list $5 / $30, so a page that does not carry a verification date is no longer safe to read.

Terra and Luna did not change. The promotional footnote is attached to gpt-5.6-sol alone. Whatever happens in November is confined to one model.

The 20% headline is the floor, not the saving

Input fell 20% and output fell 33%. Almost every report of this change led with the smaller number. Which one you actually get depends entirely on how output-heavy your workload is, and the gap between the two is large.

Our arithmetic. For a workload that generates r output tokens per input token, the old cost per million input tokens was 5 + 30r and the new cost is 4 + 20r. The saving is therefore (1 + 10r) / (5 + 30r):

Workload shape (output tokens per input token)	Typical of	Old cost	New cost	You save
0 — output negligible	Classification, extraction, scoring	$5.00	$4.00	20.00%
0.1	Long-document summarisation	$8.00	$6.00	25.00%
0.2	RAG answering over retrieved context	$11.00	$8.00	27.27%
0.5	Agent turns with moderate tool output	$20.00	$14.00	30.00%
1.0	Balanced chat	$35.00	$24.00	31.43%
2.0	Code generation, long drafting	$65.00	$44.00	32.31%
∞ — input negligible	Short-prompt generation	—	—	33.33%

Costs are per million input tokens plus the corresponding output. The practical range is 20% to 33%, and the 20% end is reached only by workloads that barely generate anything. Most real traffic lands between 27% and 32%.

The cut changed Sol’s shape, not just its level

This is the part no summary of the announcement mentions, and it is visible in one division.

Our arithmetic on the published standard-tier rates:

Model	Input	Output	Output as a multiple of input
GPT-5.6 Sol — before 21 August	$5.00	$30.00	6×
GPT-5.6 Sol — now	$4.00	$20.00	5×
GPT-5.6 Terra	$2.00	$12.00	6×
GPT-5.6 Luna	$0.20	$1.20	6×

All three GPT-5.6 models priced output at exactly six times input. Sol is now the only one that does not. Because output was cut harder than input, the cut is not a uniform discount — it deliberately rebalances the model toward generation-heavy work, and it is what produces the 20-to-33 spread above.

It also gives the November question a shape. Our inference, not a vendor statement: if the promotional pricing simply lapses back to $5 / $30, Sol returns to the 6× multiple its siblings never left. OpenAI has published no revert price, so treat that as a plausible reading of the rate card’s symmetry and nothing more.

The full rate card, all four tiers

The pricing page publishes four tables for the flagship models — Standard, Batch, Flex and Fast mode — each with eight rate columns. These are the published figures for GPT-5.6 Sol, per million tokens, as of 23 August 2026:

Tier	Input	Cached input	Cache write	Output	Input >272K	Cached >272K	Write >272K	Output >272K
Standard	$4.00	$0.40	$5.00	$20.00	$8.00	$0.80	$10.00	$30.00
Batch	$2.00	$0.20	$2.50	$10.00	$4.00	$0.40	$5.00	$15.00
Flex	$2.00	$0.20	$2.50	$10.00	$4.00	$0.40	$5.00	$15.00
Fast mode	$8.00	$0.80	$10.00	$40.00	$16.00	$1.60	$20.00	$60.00

Three things fall straight out of that table.

Batch and Flex are the same price. Not similar — identical, on all eight columns, and identical again for Terra and Luna. Both sit at exactly half the standard rate. Choosing between them is a question of how you want the request delivered, not what it costs.

Fast mode is exactly double standard on every column, which means Fast mode Sol at $8 / $40 is more expensive than Sol’s own standard price was before 21 August. Anyone who moved to Fast mode is paying 60% more for input and 33% more for output than every Sol user paid two days ago.

The whole card is generated by two numbers and three multipliers. The documentation states two of them: “Cache writes are billed at 1.25x the uncached input token rate” and “Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request.” The third — cached input at one-tenth of input — is not stated in those terms but reproduces every cached-input figure on the page exactly. Applying all three to the input and output rate of any tier regenerates all ninety-six published figures across the three models without a single discrepancy. That internal consistency is the strongest available check that the table has been read correctly.

Where the old and new rates sit line by line

Our arithmetic. OpenAI publishes only the current card, so the “before” column below is derived: the pre-cut $5 / $30 base rates with the same three multipliers applied. It is not a dated capture of the old page.

Standard-tier line	Before 21 Aug (derived)	Now (published)	Change
Input	$5.00	$4.00	−20%
Cached input	$0.50	$0.40	−20%
Cache write	$6.25	$5.00	−20%
Output	$30.00	$20.00	−33.33%
Input >272K	$10.00	$8.00	−20%
Output >272K	$45.00	$30.00	−33.33%

Every input-derived line falls by exactly 20% and every output line by exactly one third, because all of them are fixed multiples of the two base rates. One number is worth pausing on: a cache write now costs $5.00 per million tokens — precisely what an ordinary input token cost before 21 August.

The 272K cliff is worth more than the price cut

The long-context rule is not a tier and not a taper. As published: the doubled input and 1.5× output rates apply “for the full request”, and the trigger is the input token count alone.

So it is a cliff, and it is a steep one. Our worked example on standard-tier Sol, a prompt with 2,000 output tokens:

Input tokens	Input cost	Output cost	Total
272,000	$1.0880	$0.0400	$1.1280
272,001	$2.1760	$0.0600	$2.2360

One additional input token adds $1.11 to the bill and very nearly doubles it — a factor of 1.98. On Fast mode the same single token costs $2.22.

Set that against the price cut itself, which is worth 20–33%. Managing the 272K boundary is worth roughly five times more than the discount, and unlike the discount it is entirely within your control. It also does not expire in November.

The scale of the exposure is easy to miss: the model page gives GPT-5.6 Sol a context window of 1,050,000 tokens and a maximum output of 128,000. The threshold sits at 272,000. Our arithmetic: 272,000 of 1,050,000 is 25.9%, so roughly three-quarters of the usable context window is priced at the doubled rate. The advertised million-token window and the cheap rate card do not overlap for most of their range.

Prompt caching: the ratios did not move

Cache reads cost one-tenth of input and cache writes 1.25× input, on every tier and every model. Both are defined as multiples of the input rate, so the price cut moved them in lockstep and changed no cache decision at all.

Our arithmetic on the break-even, using standard-tier Sol. Sending the same prefix twice without caching costs 2.00× the input rate. Writing it once and reading it once costs 1.25 + 0.10 = 1.35×. So caching pays for itself on the very first reuse, saving 32.5% on two sends, and the saving approaches 90% as reuse grows. In dollars on Sol: $8.00 versus $5.40 for two sends of a million-token prefix.

That ratio is unchanged by the cut, and it will be unchanged by whatever happens in November — which makes it the more durable lever of the two.

Questions this raises

Do I have to change anything to get the new price?

Nothing in the documentation asks you to. The rate is attached to the existing gpt-5.6-sol model ID; no new snapshot or parameter is published alongside the change.

What happens on 22 November?

Not stated. The only published commitment is “at least through November 21, 2026”, which is a floor on the duration and not a schedule. OpenAI has published neither an end date nor a revert price. If your unit economics only work at $4 / $20, that is a risk to carry deliberately rather than an assumption to build on.

Does the cut apply to Terra and Luna?

No. Terra stands at $2 / $12 and Luna at $0.20 / $1.20, and the promotional footnote names Sol only. Sol at $4 / $20 is now twice Terra’s input price rather than 2.5 times it, so the case for stepping down to Terra is weaker than it was on 20 August.

What about the Ultrafast tier?

Announced on 13 August 2026 as a new API service tier for GPT-5.6 Sol, described in the changelog as up to 14× faster than standard processing and in limited preview. It has no rates on the pricing page, which publishes Standard, Batch, Flex and Fast mode only. We could not find a published Ultrafast rate anywhere.

Is a request classified as long-context by its prompt or its total tokens?

By the prompt. The rule is written against “prompts with >272K input tokens”; generated output does not push a request over the threshold, though it is then billed at the higher output rate once the prompt has.

Do gateway and reseller prices track this?

Not reliably. Third-party gateways publish their own rate cards and run their own separate promotions on different end dates. Several were advertising discounts on GPT-5.6 Sol in the days around this change that are theirs, not OpenAI’s. Compare against the vendor page before assuming a quoted rate is list price.

What we could not verify

• Actual billed rates. We hold no OpenAI account and have not seen an invoice. Everything here is the published rate card and arithmetic on it, not observed billing.
• The pre-cut card beyond the two base rates. The $5 / $30 figures are pinned by OpenAI’s own stated 20% and 33% reductions and corroborated by wire coverage. Every other “before” figure in this article is our derivation using the current multipliers, not a dated capture of the old page, and it assumes those multipliers were the same before 21 August. The documentation states them as current rules without dating them.
• Whether the cached-input ratio is a published rule. One-tenth of input reproduces every cached-input figure on the page exactly, but we did not find it stated as a rule the way the 1.25× and 2×/1.5× multipliers are.
• What happens after 21 November 2026. No revert price and no end date are published, only the “at least through” floor.
• Ultrafast tier pricing. The tier is announced and in limited preview; no rate is published.
• Whether ChatGPT Work and Codex credits fell by the same percentages. Wire coverage of the 21 August announcement mentions credit pricing alongside API pricing. The changelog entry we read is API-scoped, and we found no published credit rate card to check the percentages against.
• Boundary behaviour at exactly 272,000 tokens. The rule reads “>272K”, which we take as strictly greater. Whether OpenAI’s counter treats 272,000 as 272K exactly, and how it rounds, is not documented.

Sources and method

Compiled 23 August 2026 from OpenAI’s official developer documentation: the GPT-5.6 Sol model page, the API pricing page (flagship-model Standard, Batch, Flex and Fast mode tables), and the API changelog entry dated 21 August 2026. The price sentence, the promotional-pricing sentence, the cache-write multiplier and the >272K rule are quoted verbatim as they stood on that date. All ninety-six rate figures were read from the four published tables and cross-checked against the documented multipliers, which reproduce them without discrepancy. All percentages, ratios, break-evens, worked examples and “before” derivations are our own arithmetic on those published numbers and are labelled where they appear. We hold no OpenAI API account and have not run these models. Prices and billing rules change — and this one carries an explicit expiry floor — so verify against the vendor before you commit.