Home › Blog › OpenAI pricing page bug fixed

The OpenAI Pricing Bug We Left Open Last Time Is Fixed

22 November 2026 · AI Model Pricing · 5 min read

Two days ago, while backfilling three new AI models into the calculator, I checked our Anthropic pricing page and found it quoting two retired Claude models at prices 25% too high. I fixed that page the same day and, in the same post, flagged that a quick check of the OpenAI page turned up the identical bug — still open, not yet fixed. Today I went back and closed it.

What was actually on the page

FieldBefore (Nov 20)After (Nov 22)
Flagship rowGPT-6 Astra — $10.00 / $50.00GPT-6.2 Astra — $8.00 / $40.00
Mid-tier "new" rowGPT-6.1 Sol — $2.00 / $10.00GPT-6.2 Sol — $2.00 / $10.00 (unchanged)
Promo note"through at least November 21, 2026""still active per our pricing registry"
Last verified2026-11-152026-11-22

GPT-6.2 Astra and GPT-6.2 Sol both launched November 17 — five days before I got to the page. The flagship row kept showing GPT-6 Astra at the pre-cut $10.00/$50.00 rate, overstating the real cost of running the current flagship by 25% on both input and output, exactly the same ratio and exactly the same bug as the Claude page had. There was no row for either new model at all; someone reading the page on November 20, 21 or 22 would have had no way to know GPT-6.2 existed.

On a workload burning 10M input and 1M output tokens a month, the page was quietly telling readers to budget $150.00 for a bill that's actually $120.00 — a $30/month gap that compounds across every reader who priced a project off that specific number.

A second, smaller bug in the same paragraph

While I was in there I noticed the GPT-5.6 Sol promo row said the $4.00/$20.00 discount rate was valid "through at least November 21, 2026." Today is November 22. That date had already passed. I don't have evidence the rate actually changed — our pricing registry (models.js, the same file the live calculator reads from) still lists GPT-5.6 Sol at $4.00/$20.00 — so I didn't guess at a new expiry. I replaced the hard date with a statement that ties the claim to something that's actually checked rather than a date that silently goes stale the moment it passes. Hardcoded "through at least [date]" language is a trap: it reads as current right up until the date in the text is in the past, and nothing on the page tells you that happened.

What this is (still) about

Same root cause as two days ago: the dedicated vendor pages — /openai.html, /anthropic.html and the rest — are hand-written HTML, built separately from the model registry the calculator actually reads from. A new model launching and a vendor page not picking it up for 5 days is a lag I can live with. A page sitting on a pre-cut rate with no new-model row at all until someone happens to re-read it, or a promo note with a date that quietly expired, is the part worth catching earlier than "tripped over it while fixing something else" — twice in one week, on two different vendor pages.

We've now checked and fixed two of roughly a dozen dedicated vendor pages after finding the same pattern on both. The rest haven't been individually re-verified against the current registry yet — that's the honest state, not a claim that every page on the site is current.

Price GPT-6.2 Astra on the GPT-6 Astra API calculadora de costes, see every current model side by side on the Comparación de precios de modelos de IA, or run a full workload on the Calculadora de costos de API de IA.

Model prices sourced from our own pricing registry (models.js), cross-checked against each model's announcement date, 2026-11-22. The English openai.html page is fixed; translated copies are re-synced from it the same day. Prices change — confirm current rates on the provider's own pricing page before budgeting.