4 of the 5 Groq models we tracked are gone, on a median of 28 days' notice
Groq's deprecations page, read on 26 October 2026, states 13 announcements with both a notice date and a shutdown date, a median of 28 days apart. 4 of the 5 Groq models on our own audit panel have been retired since June. Below, the dates Groq states beside the dates our probes observed, how long a named replacement lasted, and why an older model is not always a dead one.
In this article
What Groq says, and what we saw
We read Groq's deprecations page on 26 October 2026, saved a copy with a hash, and parsed its tables. Separately, we have been probing a small panel of models for our own AI recommendation study since June. This page sets the dates Groq states beside the dates our probes observed.
Of the 5 Groq-hosted models on our own audit panel, 4 now return model_not_found and are absent from Groq's catalogue. Our roster, not a market rate.
Across 13 announcements that state a date (range 7 to 91). OpenAI's median on its page is 182 days, Anthropic's 63.
Median time from Groq naming a model as the recommended replacement to that replacement being retired itself, across 7 replacement models. The shortest was 89 days.
Both superseded Gemini models on our panel are still served and listed, and Google's page states no shutdown date for either.
4 models, stated against observed
A death date is an interval: the last capture that answered and the first probe that failed. We never saw a model die on a particular day.
| Model | Groq announced | Groq stated shutdown | Notice | Our last answered capture | First confirmed dead | What we can say |
|---|---|---|---|---|---|---|
| llama-3.3-70b-versatile | 17 June 2026 | 16 August 2026 | 60 days | 17 August 2026 | 5 September 2026 | Last answered 1 day after the stated shutdown date |
| meta-llama/llama-4-scout-17b-16e-instruct | 17 June 2026 | 17 July 2026 | 30 days | 8 July 2026 | 5 September 2026 | Consistent with the stated date |
| qwen/qwen3.6-27b | not stated | 14 September 2026 | no figure | 1 September 2026 | 23 September 2026 | Consistent with the stated date Our own records disagree by six days on the last answer: a partial 7 September capture file may also have answered. |
| groq/compound | 24 August 2026 | 21 September 2026 | 28 days | none | 12 October 2026 | Never captured by us; first failed probe after the stated date |
Still alive on Groq as of 26 October 2026: openai/gpt-oss-120b, the model Groq most often names as a replacement. Groq's catalogue held 14 models on 5 September and 11 on 26 October.
How much notice, next to the other providers
Groq's median notice is 28 days. The same calculation on the OpenAI and Anthropic pages gives 182 and 63 (see that comparison).
Median days between telling developers and shutdown
Groq, most recent announcements with a stated date (by announcement date)
How long a named replacement lasted
Each retirement on Groq's page names a replacement. We followed those names forward: 7 of the replacement models Groq has named were later retired themselves. The bars count from the announcement that last named each one to its own shutdown date.
Where a page states no announcement date for the retirement that named the replacement, the count starts from that retirement's shutdown date instead. Both bases are recorded for every row in mortality-ledger.json. The pairs are not independent: several retirements name the same replacement.
Superseded is not dead
Not every older model on our panel has been retired. gemini-2.5-flash and gemini-2.5-flash-lite are both still served and listed on Google's catalogue, newer Gemini models are listed beside them, and Google's deprecations page, read on the same day, states "No shutdown date announced" for both. We label such a model superseded and keep probing it. A study that drops an older model the day a newer one ships would count a live model as dead.
How we know, and what we did not verify
Stated dates, announcement dates and replacements come from the HTML tables of the page below. The build stops if the tables stop matching. Observed dates come from our own weekly captures and two catalogue reads.
| Page | Date | SHA-256 (first 16) |
|---|---|---|
| console.groq.com/docs/deprecations | 26 October 2026 | c42c90e75fa41740 |
- One provider states the dates (Groq) and one small panel was observed: 10 models, 5 of them on Groq. This is a case study, not a market rate.
- 3 panel rows (Cohere, ChatGPT web, Perplexity web) are not checked: no model catalogue was queried for them.
- Groq's page was parsed for all 30 shutdown rows, not only the panel models. Notice figures exist only where the page states an announcement date.
- Our reliability study on lucreya.com counts a different roster: 9 models including groq/compound-mini and gpt-oss-20b, 5 of them retired. The two agree on the retirement status of every model they share; they differ in which models they cover.
- We have not asked Groq to comment before publication. A provider or reader who finds a wrong figure can tell us; we correct it and date the correction.
Part of the series: retirement notice periods, price calendar, Gemini API pricing. Data: mortality-ledger.json in the AI Price Watch dataset.
FAQ
How much notice does Groq give before shutting a model down?
On Groq's deprecations page, read on 26 October 2026, 13 announcements state both a date developers were told and a shutdown date. The median gap is 28 days, ranging from 7 to 91. The page says the shutdowns apply to free and developer tier usage and that enterprise customers with a committed-spend contract are not affected.
Which models on your panel did Groq retire?
llama-3.3-70b-versatile, meta-llama/llama-4-scout-17b-16e-instruct, qwen/qwen3.6-27b, groq/compound. For each, the table shows the dates Groq states beside the last date our probes got an answer and the first date they got model_not_found. We never captured groq/compound, which was added late as an optional grounded leg.
Did any model answer after its stated shutdown date?
One of our captures did. llama-3.3-70b-versatile has a last answered capture on 17 August 2026, a day after the 16 August 2026 Groq states. We did not probe between 17 August 2026 and 5 September 2026, so we cannot say when it stopped. A time zone or a rolling shutdown could explain a one-day gap; we do not know which.
What does "a replacement is a rental" mean?
When Groq retires a model it names a replacement. We checked which of those replacements were later retired themselves: 7 distinct replacement models were, a median of 348 days after the announcement that last named them. The shortest was qwen/qwen3.6-27b, named on 17 June 2026 and retired on 14 September 2026, 89 days later.
Does a model that is superseded stop working?
Not on our evidence. gemini-2.5-flash and gemini-2.5-flash-lite are on our panel, newer Gemini models are listed beside them, and Google's deprecations page states no shutdown date announced for both. We call a model superseded, not dead, until the host says it no longer answers.