OpenAI's Pricing Page Has Four Tabs That Look the Same in Plain Text. We Published the Wrong One for Seven Days.

From September 3 to September 11, 2026 this site published the three GPT-5.6 models at half their price, with a promotion attached that did not exist as described. This is the record of that error: how two misreadings of one page combined, what caught them, how long the fix took to reach production, and the part of the fix that was itself wrong for another four days. Every price below links to the page where it can be checked, with the date it was read; the figures that come from the site’s own logs are marked as such, because those logs are not published.

What the page looks like

OpenAI’s pricing page shows the same models under four tabs: Standard, Batch, Flex and Fast mode. In a browser the tabs are obvious. In the text a script or a person extracts from the page, they are four consecutive tables with identical headers and identical model names, and nothing in the text says which is which. Read today, September 15, 2026, the four rows for GPT-5.6 Sol are $4.00 / $20.00, $2.00 / $10.00, $2.00 / $10.00 and $8.00 / $40.00 per million input and output tokens: Standard, then Batch and Flex at 50% of it, then Fast mode at 200%. The first table is Standard. On September 3 this site read the second one.

The two misreadings of September 3

The rows for GPT-5.6 Sol, Terra and Luna entered the dataset on September 3, 2026 at 17:11 UTC, in the commit that created the pricing table, with Sol at $2.00 / $10.00, Terra at $1.00 / $6.00 and Luna at $0.10 / $0.60. Those are the Batch figures. The long-context tiers above 272,000 tokens carried the same error: Sol $4.00 / $15.00, Terra $2.00 / $9.00, Luna $0.20 / $0.90, all Batch.

The second misreading was layered on top. OpenAI’s page says, above the tables, that “GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026”, and Sol’s own model page calls $4.00 / $20.00 “a 20% reduction in input pricing and a 33% reduction in output pricing”. That sentence is about Sol alone, it names no end date, only a floor, and it names no list price. The dataset turned it into a promotion on all three models that would end on November 21 with prices doubling to a “list price” of $4.00 / $20.00 for Sol, $2.00 / $12.00 for Terra and $0.20 / $1.20 for Luna. By coincidence those “list” figures were the real Standard prices, because Batch is exactly half of Standard. Terra and Luna have no promotion at all.

What caught it, and when

On September 10, 2026 the site’s price detector ran for the first time across the six first-party providers (two of Google’s pages could not be fetched in that run; a second run over Google alone at 18:01 UTC, after the script’s user agent was changed, fetched them and matched all 11 Google rows). It downloads each provider’s pricing page, hands the text to a language model together with the dataset’s rows, and asks one question per row: does the stored price match what the page says, with a verbatim quote as proof. The model never returns a price; a quote that is not found literally in the downloaded text is discarded. The method is described on the About page.

The report of that run is an internal log, not a published page; these are its figures. It was written at 17:57 UTC on September 10 and, of 35 rows checked, gave 32 matches, 3 mismatches and 0 discarded quotes. The three mismatches were Sol, Terra and Luna, each with a quote from the Standard table:

gpt-5.6-sol   | $4.00 | $0.40 | $5.00 | $20.00 | $8.00 | $0.80 | $10.00 | $30.00
gpt-5.6-terra | $2.00 | $0.20 | $2.50 | $12.00 | $4.00 | $0.40 | $5.00  | $18.00
gpt-5.6-luna  | $0.20 | $0.02 | $0.25 | $1.20  | $0.40 | $0.04 | $0.50  | $1.80

The dataset was corrected by hand the same evening: at 18:07 UTC a second run over OpenAI alone matched all six of its rows (internal log again), and at 18:09 UTC the correction was committed: Standard short-context prices, Standard long-context tiers, cache write prices, the promotion and the “list price” removed from all three rows. The correction reached production at 04:30 UTC on September 11, with the next scheduled deploy. Between detection and production, ten hours and thirty-three minutes passed, and about an hour before detection the wrong figures had been redeployed once more, with an unrelated change. Counted from the commit that introduced them to the deploy that removed them, the Batch prices were live for seven days and eleven hours.

The correction that overcorrected

Removing the promotion from all three rows was right for Terra and Luna and wrong for Sol. OpenAI’s page does state that Sol’s $4.00 / $20.00 is promotional, available at least through November 21, 2026. From the deploy of September 11 at 04:30 UTC to the one of September 15 this site showed that price with no mark, and the entry in its change log said, in so many words, that there had been no promotion to end; until the 11th it still carried the wrong promotion from the 3rd. That was found on September 15 while reconstructing the September 3 error for this post, and corrected the same day: the Sol row now carries the promotion, dated as “at least through” November 21, with the quote from OpenAI’s page and no list price, because OpenAI publishes none and this site does not compute prices. The long-context tier, $8.00 / $30.00, is defined by OpenAI as 2× input and 1.5× output of the base price for the full request, so it is promotional too and is marked as such.

What the record says now

The correction is kept, not deleted, on each of the three model pages under “Recorded price changes”, dated September 10, 2026, and it says what it is: a correction of this site’s own reading, not a change by OpenAI. It was rewritten on September 15 to add the part about Sol. A site that publishes verified prices has to publish its own errors with the same dates and the same links, or the word “verified” means nothing.

What is still not caught

The detector verifies figures, not conditions. Its answer per row is match, mismatch or not found, and its instructions tell it that when a page shows both a promotional and a list price, the line matching the dataset is a match. On September 14, 2026 it matched Sol at $4.00 / $20.00 from the Standard table while the sentence about Sol’s promotional pricing sat a few lines above that table in the very text it had been given. The sentence was in the text it received; it was not asked about it. A promotion that starts or ends without moving the headline figure will not be noticed by that check. It is a known limit, recorded in the project’s notes, and it is not fixed.

Since September 11 the detector runs every Monday. A full run over the six providers costs about $0.35 in model calls, from the run of that day, and its result goes to a person by message only when something does not match. Nothing it finds is written to the dataset by itself. On September 10 that meant a person read the report, fixed the rows by hand and left the deploy to the nightly job; since September 12 the detector also writes a proposal file for the dates it can confirm, and a person applies it with one command. The ten-hour gap above is the cost of that choice, and it is preferred to a silent overwrite.

Frequently asked questions

What are the four tabs on OpenAI's pricing page?

Standard, Batch, Flex and Fast mode. Batch and Flex are 50% of Standard and Fast mode is 200%: for GPT-5.6 Sol on September 15, 2026, $4.00 / $20.00 Standard, $2.00 / $10.00 Batch and Flex, and $8.00 / $40.00 Fast mode per million input and output tokens.

What is GPT-5.6 Sol's Standard price?

$4.00 per million input tokens and $20.00 per million output tokens, read on September 15, 2026. OpenAI says this promotional pricing is available at least through November 21, 2026, and publishes no list price for after it.

How long did AI Signal publish the wrong GPT-5.6 prices?

Seven days and eleven hours: the Batch prices entered with the commit of September 3, 2026 at 17:11 UTC and were removed by the deploy of September 11 at 04:30 UTC. The price detector found the mismatch on September 10, 2026.

Does the weekly price detector catch promotions?

No. It verifies figures, not conditions, so a promotion that starts or ends without moving the headline price is not noticed; that limit is known and not fixed.