If your information was crawled by a model and became part of its knowledge today, it can take a year or more for that information to be shown in the next model release.
This is the basic 3-step process:
Step : 1 (approximately 6 months):
Information is gathered by web-crawlers in what is called “an initial unsupervised pre-training of knowledge”. The machine gathers massive amounts of data from the internet and public databases to learn general language patterns, and this is where the model learns about you, your brand, and your company.
Step : 2 (approximately 6 months):
Once that base knowledge is gathered, the machine has to be trained on how to answer questions and interact safely via chat. This is called a “fine-tuning” stage (there can be more than one of these to make responses better and better).
Step : 3
The models are released to the public.
There is a time lag between Steps 1, 2, and 3. Below are some examples of the lag between the knowledge cutoff date (end of Step 1) and the public release date (end of Step 2):
| Model | Knowledge‑cutoff date (given) | Public announcement / release date | Lag (months)¹ |
|---|---|---|---|
| ChatGPT‑4o (June‑cutoff update) | June 2024 | 29 Jan 2025 (help.openai.com) | 7 |
| Claude Haiku 3.5 | July 2024 | 22 Oct 2024 (anthropic.com) | 3 |
| Claude Opus 4 | March 2025 | 22 May 2025 (en.wikipedia.org) | 2 |
| Claude Sonnet 3.7 | Nov 2024 | 24 Feb 2025 (en.wikipedia.org) | 3 |
| laude Sonnet 4 | March 2025 | 22 May 2025 (en.wikipedia.org) | 2 |
| DeepSeek R1 Distill Llama 70B | May 2024 | 20 Jan 2025 (llm-stats.com) | 8 |
| GPT‑4.1 | June 2024 | 14 Apr 2025 (techtarget.com) | 10 |
| GPT‑4.1 mini | June 2024 | 14 Apr 2025 (docsbot.ai) | 10 |
| GPT‑4.1 nano | June 2024 | 14 Apr 2025 (docsbot.ai) | 10 |
| GPT‑4o(original) | Oct 2023 | 13 May 2024 (openai.com) | 7 |
| GPT‑4o mini | Oct 2023 | 18 Jul 2024 (en.wikipedia.org) | 9 |
| Gemini 2.0 Flash | Aug 2024 | 05 Feb 2025 (cloud.google.com) | 6 |
| Gemini 2.0 Flash Lite | Aug 2024 | 25 Feb 2025 (cloud.google.com) | 6 |
| Gemini 2.5 Flash Preview 05‑20 | Jan 2025 | 20 May 2025 (console.cloud.google.com) | 4 |
| Gemini 2.5 Pro Preview 06‑05 | Jan 2025 | 05 Jun 2025 (cloud.google.com) | 5 |
| Grok 3 | 17 Nov 2024 | 17 Feb 2025 (en.wikipedia.org) | 3 |
| Grok 3 Mini | 17 Nov 2024 | 17 Feb 2025 (artificialanalysis.ai) | 3 |
| Meta Llama 3.3 | Dec 2023 | 06 Dec 2024 (huggingface.co) | 12 |
| Meta Llama 4 Maveric | Aug 2024 | 05 Apr 2025 (huggingface.co) | 8 |
| Meta Llama 4 Scout | Aug 2024 | 05 Apr 2025 (huggingface.co) | 8 |
| Perplexity Sonar (web‑search) | Unknown | 21 Jan 2025 (techcrunch.com) | N/A |
| Perplexity Sonar Pro (web‑search) | Unknown | 21 Jan 2025 (artificialanalysis.ai) | N/A |
| Perplexity R1‑1776 (no web‑search) | Unknown | 18 Feb 2025 (perplexity.ai) | N/A |
| Qwen3 32B | 28 Nov 2024 | 29 Apr 2025 (llm-stats.com) | 5 |
The average lag time is just over 6 months from the end of pretraining to public release. Adding in the pretraining time of approximately 6 months, it will take approximately 1 year for new information to show in the next generation of large language models.