GPT-5.6 Sol Cuts Prices 20%? OpenAI's Own Docs Say It's Promotional — Guaranteed Only Through Nov 21 [2026 Verified]

On 2026-08-22, a Hacker News headline climbed the front page: “GPT 5.6 Sol 20% price reduction.”

We clicked through to the OpenAI official model documentation it pointed at. The sentence is there, word for word:

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

(In the quotation above, the source writes the amounts with a bare $; this site prefixes all amounts as US$ per our display convention. The figures are unchanged.)

But here is the very next sentence in the same paragraph:

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

In plain terms: this is promotional pricing, and OpenAI only commits to it “at least through” November 21, 2026. What happens after that date is not stated in the docs.

And once you open the full pricing tables, there are three surcharges that can wipe out that 20% entirely. Every number below comes from pages we read directly.

I. What Happened, in 30 Seconds

• This is an API price change, not a ChatGPT subscription change. Your ChatGPT monthly fee is unchanged (our ChatGPT tool page records the Taiwan list prices NT$270 / NT$690 / NT$3,300, verified 2026-08-09). This article is about developer API billing.
• New pricing: per 1M tokens, input US$4.00, output US$20.00, cached input US$0.40.
• Versus GPT-5.5: input US$5.00, output US$30.00. So the 20% / 33% figures are measured against 5.5.
• The gpt-5.6 alias routes to Sol — the most expensive model in the family.
• Specs: 1,050,000 token context window (maximum input 922,000), 128,000 max output tokens, knowledge cutoff 2026-02-16.
• The API free tier cannot use this model: on the official rate-limit table, the Free row reads “Not supported.”

(All of the above comes from OpenAI’s official model documentation and official pricing page, read directly by this site on 2026-08-22. Links at the end.)

II. The Pricing Table, Laid Out

Per 1M tokens, in USD, from the official pricing page’s “Standard pricing data” table (verified 2026-08-22):

One easily missed fact first: Sol is cheaper than 5.5, but more expensive than 5.4. GPT-5.4 is US$2.50 input and US$15.00 output — both below Sol. The “price cut” is relative to the previous frontier model, not to the GPT-5 family as a whole.

The same family also has cheaper siblings: terra (US$2 / US$12) and luna (US$0.20 / US$1.20). If your workload does not genuinely need frontier reasoning, switching model tiers saves far more than this promotion does.

III. Three Surcharges That Eat the Discount

This is the core of the article. OpenAI puts all three in fine print under the price — and each one hits harder than 20%.

1. Above 272K context → 2x input, 1.5x output, charged on the whole request

Official wording: “Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request.”

Note for the full request: it is not just the overspill that costs more — the entire request is billed at the higher rate. The pricing page spells the numbers out in its “Long context” columns:

See it? Once your prompt crosses 272K, Sol’s output price is US$30 — identical to GPT-5.5. The “33% output reduction” is fully cancelled in that scenario, and input is now 60% more expensive than 5.5 (US$8 vs US$5).

Sol’s headline feature is its 1.05M token context window. But actually filling that window puts you in the 2x billing tier. Those two facts are attached to each other.

2. Cache writes are billed separately at 1.25x

Official wording: “Cache writes are billed at 1.25x the uncached input token rate.” On the pricing table, Sol’s cache write rate is US$5.00 (= US$4 × 1.25).

Worth noting: on that same table, the “cache writes” cells for GPT-5.5 and GPT-5.4 are both “–” (no figure listed), while Sol has an explicit US$5.00. We report the table as it stands and do not speculate about why.

Cached reads really are cheap (US$0.40, just 10% of the input rate) — but you pay the US$5.00 write first. A cache has to be hit enough times to pay for itself; a one-shot long prompt costs you more, not less.

3. Fast mode is 2x standard

The pricing page also notes that Priority processing was renamed Fast mode on 2026-07-30; service_tier accepts either "priority" or "fast".

One more surcharge: regional processing (data residency) endpoints carry a 10% uplift for eligible models released on or after 2026-03-05.

IV. Turned Into Numbers You Can Feel

Say you burn 500K input tokens plus 100K output tokens per day, over 30 days (= 15M input, 3M output):

At roughly 1 USD ≈ 31.8 TWD (2026-08-22, two rate sources listed at the end), the Sol promotional figure is about NT$3,816/month, saving roughly NT$1,431/month against GPT-5.5’s ~NT$5,247.

But that saving has an expiry date. OpenAI commits only through 2026-11-21 — about three months from today.

V. Our Read

This is not the first time. The three recent “price cuts” we have checked follow nearly the same shape:

• Gemini 3.7 Flash claimed half price — a promo, doubling on 2027-01-01.
• Claude Sonnet 5 cancelled its price increase — the opposite case: a permanent standard price, stated as such by the vendor.
• Runway, Kling and Pika all “cut prices” the same day — four companies, four different illusions, not one real cut.

The test is a single question: does the vendor write “permanent / standard,” or “promotional / through ”? Sol says the latter — and it uses “at least through,” which makes even the expiry date a floor rather than a fixed point.

AMO (the one who hunts for traps): “The 1.05M context window and the 272K surcharge threshold are on the same page — one in the headline, one in the fine print.”

PIMI (the one who looks for the upside): “But saving 27% is real, and the Batch half-price is officially documented. Three months is enough to finish a project cycle.”

Both cats are right. Use it now — just don’t budget at US$4 / US$20.

VI. Three Things You Can Do Today

1. Put 2026-11-21 in your calendar. OpenAI has made no commitment past that date, so for annual budgeting use GPT-5.5’s US$5 / US$30 as your ceiling. (That is a conservative estimate anchored to the only official figure available — it is not an announced post-promo price.)
2. Check whether your prompts exceed 272K. If they do, the whole request bills at 2x input and 1.5x output. Prompt compression or retrieval will save you more than switching models will.
3. Move anything that can wait to Batch. The official table lists Batch at half the standard rate (US$2 / US$10) — the only way to halve your bill without changing model.

Sources

Official (each read directly by this site on 2026-08-22)

• OpenAI official model documentation, “GPT-5.6 Sol” (US$4/US$20, the 20%/33% reduction, promotional pricing at least through 2026-11-21, >272K surcharge, 1.25x cache writes, 1.05M context, Free tier “Not supported”): https://developers.openai.com/api/docs/models/gpt-5.6-sol | verified 2026-08-22 (HTTP 200; we also read the official Markdown version of the same page via the .md suffix and cross-checked — both versions agree)
• OpenAI official API pricing page (Standard / Batch / Fast mode tables, long- and short-context columns, cache-write column, Fast mode rename date, 10% data-residency uplift): https://developers.openai.com/api/docs/pricing | verified 2026-08-22 (HTTP 200)

News trigger

• Hacker News thread, “GPT 5.6 Sol 20% price reduction” (2026-08-22, links to the official doc above): https://news.ycombinator.com/item?id=49396590 | used solely as a topic trigger; every figure in this article was read from the official pages, and no thread content is quoted.

Second-hand (disclosed per our rule 9.7)

• Exchange rate (1 USD ≈ 31.8 TWD, 2026-08-22): https://open.er-api.com/v6/latest/USD (31.812, last updated 2026-08-22 00:02 UTC) | cross-referenced against https://cdn.jsdelivr.net/npm/@fawazahmed0/currency-api@latest/v1/currencies/usd.json (2026-08-22, 31.835)
  • The exchange rate is not from OpenAI’s pricing page; TWD figures are this site’s estimate. OpenAI bills in USD, and your actual charge depends on your payment provider’s rate.

Site data

• The Taiwan subscription list prices on our ChatGPT tool page (NT$270 / 690 / 3,300) come from a 2026-08-09 check reading the official Traditional Chinese site from a Taiwan connection, not from this article’s date. Those figures are unrelated to the API pricing here and are included only to distinguish “the subscription fee did not change.”

Cross-verification status (disclosed per rules 9.1 / 9.2): the US$4 / US$0.40 / US$5 / US$20 figures, the 20% / 33% reduction, the 2026-11-21 promotional floor, the >272K surcharge and the 1.25x cache-write rate were all read consistently from three places: OpenAI’s official model page (HTML), the official Markdown version of that same page, and the official API pricing page. All three sit on OpenAI’s own domain, and we have not yet obtained word-for-word cross-confirmation from an independent third-party outlet — so per rule 9.2 this is flagged as “single-organisation official source, consistent across multiple pages.” The vendor’s own announcements govern.