GPT-5.6 Sol Drops to $4/$20 on the API

Promotional rates on a flagship reasoning model now sit at $4 per million input tokens and $20 per million output tokens, according to OpenAI's live pricing table. Cached input is $0.40. The page says the promotion lasts at least through November 21, 2026. Reuters reported the same day that the cut is "more than 20%" for the next three months, and that it applies to the API plus eligible ChatGPT Work and Codex credits.

ChatGPT Plus, Pro, and Business subscription prices do not move. If you only talk to ChatGPT in the app, nothing on your invoice changed. The people who feel this are API callers and anyone whose Work or Codex usage is billed in tokens or credits.

"As we continue to push the frontier of capabilities while improving efficiency, we're dropping API and credit pricing of GPT-5.6 Sol by over 20% for the next 3 months."

That line is from OpenAI's developer-community announcement on August 21, 2026. The same post says Pro, Plus, and Business subscription usage remains unchanged, and that the new rates are rolling out across eligible Work and Codex credit plans.

$5/$30 became $4/$20 on short context

Reuters published the before-and-after against the same table: previous standard short-context prices were $5 input and $30 output. The new pair is $4 and $20. Input is a 20% cut. Output is about 33%. Cached input went from $0.50 to $0.40.

The official table also lists long-context Sol at $8 input and $30 output (cached input $0.80). Batch and Flex both list Sol short-context at $2 input and $10 output. Fast mode lists $8 input and $40 output. Regional processing endpoints add a 10% uplift on models released on or after March 5, 2026 that are eligible for data residency, per the same pricing page.

Numbers are from the OpenAI pricing table as fetched on August 23, 2026. The previous-standard row is the pair Reuters published as the rate before the August 21 cut.

Terra and Luna already moved on July 30

The community announcement includes a second table for the rest of the GPT-5.6 family. Terra is listed at $2 input / $12 output (was $2.50 / $15). Luna is listed at $0.20 / $1.20 (was $1 / $6). Those two cuts landed on July 30, 2026, three weeks before Sol. Reuters described them as a 20% Terra cut and an 80% Luna cut.

Sol was the holdout. For three weeks the flagship stayed at $5/$30 while the cheaper siblings dropped. The August 21 post is the flagship catching up, on a clock: "at least through November 21, 2026," which is the phrasing on both the pricing page and the community thread. It is a promotion with a published floor date, not a permanent rewrite of the list price.

Reuters put Anthropic's list next to it

The same Reuters piece lists Claude Fable 5 at $10 input and $50 output per million tokens, and Claude Opus 5 at $5 input and $25 output. Against those published lists, promo Sol is cheaper than both on input and cheaper than Fable 5 on output. Opus 5 output is still $25 against Sol's $20.

That is a list-price comparison, not a quality comparison. Token price is what you pay per million tokens on the published table. It says nothing about how many tokens a given job burns, or whether a cheaper model finishes the job. People shopping "the frontier API" by the spreadsheet will see Sol undercut both Claude rows for the length of the promotion. People shopping by evals still have to look at evals.

Reuters framed the cut as a response to "growing competition from Anthropic and Chinese AI models." OpenAI's own post frames it as efficiency: "push the frontier of capabilities while improving efficiency." Both sentences are in the public record. Neither is a benchmark.

Who actually pays less

Three groups show up in the announcement:

• Direct API callers using the gpt-5.6-sol model id on standard, batch, flex, or fast tiers.
• ChatGPT Work users on eligible plans whose usage is drawn from credits, as those credits buy more tokens at the new rate.
• Codex users on eligible credit plans, same mechanic.

Three groups do not:

• Plus, Pro, and Business subscribers whose usage is bundled into the monthly fee.
• Anyone still on GPT-5.5 or older ids. Those rows on the pricing table did not move with this post.
• Anyone calling Sol through a third-party proxy whose markup is not tied to OpenAI's list. Check the proxy.

A rough back-of-envelope from the published numbers: one million input tokens plus one million output tokens on standard short context used to cost $35. It now costs $24. That is $11 per that pair, about 31% off the combined bill, weighted toward output. Jobs that emit a lot of tokens (long agent traces, verbose code, multi-step tools) save more in dollars than jobs that read a lot and write a little.

How to point a call at the new rate

The model id on the pricing table is gpt-5.6-sol. There is no new id for the promotion. The same string now bills at the promo row. If your client already sends that id, you do not change code to get the new rate. You will see it on the invoice.

If you have been routing cheap traffic to Luna and expensive traffic to Sol, the spread shrank. Luna is still far cheaper ($0.20/$1.20 vs $4/$20). Terra sits in the middle ($2/$12). The reason to pick Sol is whatever you already believed about its quality, not a new capability drop in this announcement. The announcement is a price line.

Batch and Flex halve the standard promo ($2/$10). Fast mode doubles it ($8/$40). Pick the service tier the way you already do; only the dollars attached to each tier moved.

Daybreak aliases are on the same page: daybreak-blue-latest currently points at gpt-5.6-sol and inherits that row. If you pinned the alias instead of the dated id, you are on the promo for as long as the alias keeps pointing here, and you will follow it when OpenAI retargets the alias.

The useful question for a newcomer is not "is Sol worth it now." It is "am I paying per token at all." If the answer is no, because you live in ChatGPT Plus, this post is a spectator note. If the answer is yes, the live table is the source, the date on the promotion is November 21, 2026 unless OpenAI extends it, and the siblings on the same family (Terra, Luna) are the actual alternatives when the job does not need the flagship row.