ChatGPT knowledge cutoff: stated vs measured in 17 AI models

What is ChatGPT's knowledge cutoff?

Your ChatGPT cutoff depends on your plan and on what you pick in the model menu, because OpenAI dates each model separately on its model pages. A free account knows about two months less than a paid one. GPT-6 Pro reaches 10 days further than standard GPT-6. Here's the model and the cutoff behind each part of ChatGPT, as of October 2026:

OpenAI's help article on GPT-6 in ChatGPT spells out the split: "When GPT-6 is selected, GPT-6 Sol powers Instant, Medium, High, and Extra High". Plus only gets Medium and High. The Pro level is the one that uses GPT-6 Astra. GPT-6 is still rolling out, and OpenAI says a paid account that doesn't see it in the menu may not have it yet. OpenAI publishes just one date for each model, with no separate training data cutoff. A model that you pick by hand keeps its own date, so if you choose GPT-5.6 Sol, you're back in February.

We checked whether the April date holds. On Whizi's Knowledge Cutoff Bench, run on 15 October 2026 with web search off, GPT-6 Sol's memory reached April 2026. That's exactly what OpenAI says. Only one other model out of ten did the same.

Do AI models really know everything up to their stated cutoff?

No. In the 15 October 2026 run of Whizi's Knowledge Cutoff Bench, 8 of the 10 models whose labs publish a cutoff remembered less than the lab claims they do. Among those 8, the median gap was 4.5 months. Gemini 3.5 Flash-Lite was the furthest off, at 13 months, and Gemini 3.8 Flash came next at 12.

All 17 models answered the same 132 questions, 4 for each month from January 2024 to September 2026. Each question is about something an official record dates to that month. Web search was off and the prompt gave no date, so a model could only answer from training. Its measured cutoff is the last month where it got at least half right, averaged with the two months before, so one lucky guess can't move it. GPT-6 Sol only just made April. It got 1 of 4 April questions right, after 4 of 4 in February.

Google's own cards hint at the gap. The Gemini 3.8 Flash model card gives March 2026 as the cutoff, then warns that in some domains the model's knowledge is "limited to January 2025". The Gemini 3.5 Flash-Lite card carries the same warning. Our run found the older date was much nearer the truth. Gemini 3.8 Flash didn't get a single question right for any month after March 2025.

Which AI model has the most recent knowledge cutoff?

Claude has the newest stated cutoff. Anthropic gives June 2026 for all four of its current models: Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1. When we measured it, though, GPT-6 Sol was the newest at April 2026, a month ahead of Claude Opus 5.5. The table below lists the current models from OpenAI, Anthropic, Google and xAI. We read the stated dates from each lab's docs on 18 October 2026, and the measured ones come from the bench run on 15 October 2026, for the models that were in it.

Four labs in the test publish no cutoff at all: DeepSeek, Moonshot (Kimi), Alibaba (Qwen) and Z.ai (GLM). DeepSeek's models and pricing page lists its two API models with no date. For those seven models, a measured date is the only cutoff you'll get, since the labs don't give one. xAI is the opposite case. Its models page says "The knowledge cut-off date of Grok 4.7 is May 2026", and our run found August 2025.

What is Claude's knowledge cutoff?

Claude's current models all have a June 2026 knowledge cutoff. That's true for Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1, per Anthropic's models overview. Anthropic publishes two dates for each of them. The reliable knowledge cutoff is "the date through which the model's knowledge is most extensive and reliable", and the training data cutoff is the wider range of data it was trained on. For all four of these models, both of the dates are June 2026.

Older Claude models have earlier dates. Anthropic's help article on training data lists Claude Opus 5 at May 2026, and it puts Sonnet 5 at January 2026 and Sonnet 4.6 at August 2025.

Our test found that Claude falls short of June. Claude Opus 5.5 reached March 2026, which is 3 months early. Claude Fable 5.1 stopped at December 2025, 6 months early, even though it's the most expensive of Anthropic's four current models. Opus 5.5 still had the second freshest memory of all the models we measured. We didn't test Sonnet 5.5 or Haiku 5.5 in this run.

What is Gemini's knowledge cutoff?

Gemini 4 Argon, which is Google's newest model, has a March 2026 cutoff. Its model card was published on 13 October 2026, and it adds that "in some domains the model's knowledge may be less up to date." The card lists the Gemini app as one of the places where the model runs. Gemini 3.8 Flash and Gemini 3.5 Flash-Lite state the same March 2026 date.

Gemini 3.1 Pro states no date of its own. Its model card points to the Gemini 3 Pro card for training data, and that card says the Gemini 3 Pro cutoff "was January 2025." Google's API model pages don't show a cutoff row anymore, so on 18 October 2026 the dates were only on the DeepMind model cards.

In our run, Gemini 3.1 Pro matched its January 2025 date exactly. Gemini 3.8 Flash measured March 2025, which is a year before the date it states, and Gemini 3.5 Flash-Lite measured February 2025. We didn't test Gemini 4 Argon. Google published its card only two days before the run.

Why does ChatGPT give a different cutoff when you ask it?

ChatGPT can't look up its own cutoff. The model keeps no record of its training. Asked for the date, it predicts a likely answer from text it read, and most of that text was written about older models. Take the date from OpenAI's model page, not from the chat.

What a model does past its cutoff matters more. On questions after its measured cutoff, GPT-6 Sol said it didn't know every time instead of guessing. Claude Opus 5.5 admitted it on 96% of those questions and Grok 4.7 on 98%. DeepSeek V4 Pro admitted it on only 65%. It answered the other 35% anyway, and most of those answers were wrong.

Search fills the gap for any model. ChatGPT, Claude and Gemini each have their own web search. Whizi turns web search on by default in every new chat, for GPT, Claude and Gemini models alike. So an answer about last week comes from live pages, not from a cutoff.