11 November 2026 · Site Methodology & Corrections · 5 min read
I went looking for a new model to add to our calculator today and ended up pulling five out instead. GPT-4 Turbo, o1, o1-mini, o3-mini and o4-mini were all still sitting in our shared models.js file, each with a clean input/output price per million tokens, each feeding the dropdown on 18 different pages across this site. Every one of them is dead. Not deprecated-with-a-warning dead -- actually shut down, calls returning errors, according to OpenAI's own deprecations page.
I checked each model against OpenAI's current deprecations documentation instead of trusting what we had on file. Here is what it says, and how long each one had been dead while we kept pricing it:
| Modello | Shut down | Dead for (as of today) | What we were charging for it |
|---|---|---|---|
| o1-mini | 27 Oct 2025 | ~1 year, 15 days | $1.10 / $4.40 per 1M |
| o1 | 28 Sep 2026 | 44 days | $15.00 / $60.00 per 1M |
| GPT-4 Turbo | 23 Oct 2026 | 19 days | $10.00 / $30.00 per 1M |
| o3-mini | 23 Oct 2026 | 19 days | $1.10 / $4.40 per 1M |
| o4-mini | 23 Oct 2026 | 19 days | $1.10 / $4.40 per 1M |
o1-mini is the one that actually stings. OpenAI killed it in October of last year. We've been running a full calendar year of calculator sessions where a visitor could select it, type in a token count, and watch us hand back a confident-looking dollar figure for an API call that has not worked since 2025.
This is the same trap we fell into with Gemini fine-tuning a couple weeks back, just in a different shape. A dead model and a live one look identical in a flat pricing table: a name, an input rate, an output rate, a context window. There is no column for "does this actually respond." The only way to catch it is to stop trusting the list and go check the vendor's own page -- which is exactly what nobody had done on this file since it was built back in July.
The four October 23 shutdowns (GPT-4 Turbo, o3-mini, o4-mini, and o1-pro alongside o1 a month earlier) were part of what OpenAI itself is calling its biggest deprecation wave yet -- over two dozen legacy endpoints retired between July and October 2026. o3 and o3-pro are next, scheduled to go dark on December 11. Those two we checked and left alone, because "scheduled" isn't "dead" -- we'll pull them the day they actually are.
All five lived in one shared file, models.js, that feeds the model dropdown on the main AI cost calculator plus 17 other pages -- the Calcolatore del costo del token LLM, price comparison tool, calcolatore dei costi del chatbot, AI agent calculator and more. One bad row, eighteen pages wrong at once. That's the real lesson here: shared pricing data is efficient right up until it's stale, and then it's stale everywhere simultaneously instead of in one obvious place.
Pulled all five rows out of the base model list and regenerated the file. If you pick a reasoning model on our calculator today, the only OpenAI reasoning options are o3 and o4 -- the ones that still actually run. GPT-4 Turbo is gone; GPT-4.1, GPT-4o and the newer GPT-6 Astra/Sol line cover that tier now at a lower or comparable price anyway, so nothing of practical value was lost by removing it.
I don't have a tidy process yet that prevents this from happening again automatically -- the honest fix is just checking more often than once a quarter. For now, if you're budgeting a project against any model name, run it through our tools and double check the model is still one OpenAI, Anthropic or Google will actually answer a request on:
New to usage-based AI pricing in general? Start with the guide gratuite sui costi API. And if you want the other half of this story, see the Gemini fine-tuning post from two weeks ago -- same mistake, different provider, same fix: actually check instead of assuming the dropdown is still true.
Shutdown dates sourced from OpenAI's own API deprecations documentation, checked 11 November 2026. Prices for models still listed are reference estimates tracked in our catalog -- confirm current rates on the provider's pricing page before budgeting.