AI knowledge cutoff dates: ChatGPT, Claude, Gemini and Grok

† Confirmation level. Dates without a dagger come from the provider's current documentation. A dagger marks a value the current documentation no longer shows – in this table, dates from an earlier version of the xAI documentation – and it should be checked before you cite it.

‡ Conflicting provider sources. Two official sources from the same provider give different dates, so I don't pick one and don't calculate a gap. Claude Sonnet 4.6: the model page in Anthropic's documentation gives August 2025 (training data: January 2026), the Transparency Hub gives May 2025. Gemini 2.0 Flash: the model page in the Gemini API documentation gives August 2024, the Firebase AI Logic documentation gives June 2024.

Two dates at Anthropic. In the table I give the reliable knowledge cutoff – the date through which, according to Anthropic, the model's knowledge is most extensive and reliable. Anthropic's model documentation also publishes a separate training data cutoff, which can fall later – February and July 2025 for Haiku 4.5, January and July 2025 for Sonnet 4.5. Where the two values differ, I give the second one in the caption. The Transparency Hub describes training data in prose and, for Sonnet 4.6, gives a different cutoff than the model documentation.

*** Claude Mythos 5.1. The model does not appear in the public model catalog and has no cutoff of its own. The Transparency Hub states explicitly, however, that Mythos 5.1 and Fable 5.1 are the same model, differing only in their level of safeguards, so I assign it Fable 5.1's cutoff. This is an inference from the documentation, not a reading from a model card. The Mythos family also shows that a cutoff and availability are two different things: on 12 June 2026 Fable 5 and Mythos 5 were switched off as a result of US export controls; the controls were lifted on 30 June, and on 1 July Fable 5 returned for all users, Mythos 5 for selected US organisations. The cutoff stayed the same throughout – what changed was whether the model could be used at all.

OpenAI variant names. Sol and Luna now denote models in the GPT-6 family, but the same names appeared earlier in GPT-5.6. A variant name alone therefore indicates neither the generation nor the cutoff – the full model identifier decides, for example gpt-6-luna. On top of that, the gpt-5.6 alias routes requests to GPT-5.6 Sol.

** Gemini and two dates in one model card. Google DeepMind's model cards for 3.5 Flash-Lite, 3.6 Flash, 3.7 Flash and 3.8 Flash give a March 2026 cutoff and immediately add that in some domains the model's knowledge extends only to January 2025, in line with the policy for the Gemini 3 family. In the column I give the value declared as the main one. This is the only case in the table where the provider itself describes its cutoff as uneven – more on this in the section on the effective cutoff.

Google documentation in four places. Cutoff dates for specific Gemini models are in the model cards on deepmind.google, not in the API documentation on ai.google.dev, which declares January 2025 for the whole Gemini 3 family. The discrepancy between these two sources is only apparent: the model card gives both values, the API documentation only the old one. The third place is the Vertex AI documentation in Google Cloud, which gives one date per model – for example January 2025 for Gemini 2.5 Flash-Lite and 3.1 Flash-Lite. The fourth is the Firebase AI Logic documentation, which lists dates for older models: January 2025 for the 2.5 family and June 2024 for Gemini 2.0 Flash, whereas that model's page in the Gemini API documentation gives August 2024 (see ‡).

SpaceXAI (xAI). The company now operates under the SpaceXAI brand, while its documentation remains at docs.x.ai. Grok 4.7 (released 21 September 2026) has a May 2026 cutoff on the current model page. Grok 4.6 and 4.5 are still available in the API, but their documentation pages now redirect to Grok 4.7; the date of 1 February 2026 was given for them by an earlier version of the documentation, hence the dagger. xAI also states plainly that without search tools enabled the model has no access to current events.

Microsoft Copilot. There is no single cutoff, because a combination of models from different providers runs underneath, and queries are sometimes routed between them automatically. The date therefore depends on the product, the mode and which model answered.

The “gap to today” column. Counted from the month of the declared cutoff to the current month, in the browser, each time the page loads. For entries without a declared date I don't estimate. Where the provider gives only an approximate period, as PLLuM does (early 2025), I count from the middle of that period and mark the result as approximate.

Mistral. The model cards in Mistral's documentation give the release date, parameters, context and pricing, but not the cutoff. Compilations that assign Mistral Large 3 a “December 2025” cutoff are repeating its release date. Mistral's help centre states plainly that the company does not disclose the datasets used for training, and its model compliance page gives release dates without a cutoff.

Polish models. PLLuM gives an approximate data range in its Hugging Face model cards: early 2025 for PLLuM-12B-base-250801 and autumn 2024 for the July 2025 models. Bielik gives no cutoff in either its model cards or its technical reports. Its AI Act summary, which covers v3 11B Instruct, enters May 2025 in the official field as the latest date of data acquisition, and in the description of public datasets gives a period running to November 2025. These are data-collection dates, not a declared knowledge boundary, which is why the column says “not published”. The v3 models did, however, start training from other providers' base models: the smaller variants from Qwen 2.5, and 11B v3 from Mistral 7B v0.2. What a Polish model knows therefore also includes what its base model learned.

Qwen. The model cards for Qwen 3.6 and 3.8 describe architecture and context length, with no cutoff. In the discussion under the official Qwen3.8 repository, the only answer to a question about the cutoff was: ask the model. That is exactly the method I warn against in the section on reading cutoff dates.

DeepSeek. The model cards for V4 and V4.1 Flash describe architecture and corpus, with no cutoff. According to DeepSeek's announcement, since 14 September 2026 requests sent to deepseek-v4-pro are served by V4.1 Flash, and the deepseek-v4-flash name also points to the newer model. An API model name therefore does not settle which model actually answers.