Gemini Knowledge Cutoff Dates by Model (2026)

The Gemini knowledge cutoff is January 2025 for several documented Gemini 2.5 and 3.x models. Google has not published a universal cutoff for the whole Gemini family, including newer stable endpoints such as gemini-3.8-flash. Search grounding can also give Gemini access to newer information without changing its training cutoff.

Gemini cutoff dates by model

Google documents different Gemini knowledge cutoff dates by model, and some current models have no published cutoff. The table below covers general-purpose Gemini API models and their status as of Sept. 29, 2026.

Sources: Google’s Gemini model catalog, Gemini API release notes, deprecation schedule, Gemini 3 developer guide, Gemini 3.5 Flash guide, Gemini 3.8 Flash model page, and the Google-authored Gemini 2.5 technical report.

gemini-3.8-flash is the latest stable Flash endpoint in this snapshot. Its version number and release date do not justify assigning it the January 2025 cutoff documented for selected earlier models.

What the cutoff means

A Gemini knowledge cutoff marks the boundary of a model’s built-in, or parametric, knowledge. The date indicates how recent the information used to build that knowledge may be. It does not promise complete or accurate knowledge of every event, company, or webpage published before the cutoff.

Google’s technical report identifies pre-training dataset cutoffs of November 2023 for Gemini 1.5, June 2024 for Gemini 2.0, and January 2025 for Gemini 2.5. Google does not provide a separate cutoff date for post-training data such as instruction-tuning, preference, or tool-use data.

The cutoff also differs from:

• The model’s context-window size
• Files uploaded with a request
• Earlier messages in a conversation
• Facts included directly in a prompt
• Information retrieved through tools

A model can analyze a new document even when the document was created after its training cutoff. The document enters the request as context rather than becoming part of the model’s built-in knowledge.

The same distinction applies when comparing AI knowledge cutoff dates across major LLMs. A later date can suggest a newer training boundary, but it cannot tell you which model will retrieve, cite, or recommend your brand today.

How Gemini gets newer facts

Gemini can get facts beyond its knowledge cutoff through answer-time retrieval and information supplied with the request.

Google Search grounding

Google Search grounding connects Gemini to web content while an answer is being generated. According to Google’s Search grounding documentation, the model analyzes the prompt, decides whether Search would improve the answer, creates one or more queries, processes the results, and synthesizes a response.

Grounded responses can include the executed queries, source URLs, page titles, and citations connected to specific passages. The underlying training cutoff remains unchanged. A recent fact may have come from retrieval rather than newer training.

Grounding evidence is not guaranteed. Google Cloud’s grounding documentation says metadata can be absent when source relevance is low or available information is incomplete. An uncited answer therefore does not prove that Gemini relied only on training data.

Gemini app behavior

Gemini Apps can combine a selected model with public information and connected services. Google’s Connected Apps guidance says Gemini can use public information from Google Search, Flights, Hotels, Maps, and YouTube, plus permitted Connected Apps.

Consumer labels also hide some technical detail. Gemini Apps currently presents options such as Gemini Flash-Lite, Gemini Flash, and Gemini Pro, but those names do not map transparently to exact API IDs.

Google therefore publishes no universal cutoff for every Gemini Apps response. Results can vary with location, language, device, selected model, enabled features, and connected context.

Why Gemini answers change

Gemini answers can change even when the model’s documented knowledge cutoff stays fixed. A cutoff sets an information boundary, not a fixed response.

Prompt wording, conversation history, system instructions, cached content, tools, and function responses can all affect the output. Generation settings matter too. Google’s inference reference says temperature 0 remains capable of some variation, while a fixed random seed makes only a best effort toward repeatability.

Model routing can create larger changes. A specific stable endpoint usually points to one stable model, while the latest alias can be swapped to a newer release. Google provides two weeks’ notice before a breaking change to the model behind that alias.

On May 19, 2026, gemini-3.5-flash became the model behind gemini-flash-latest. Google also changed its default thinking effort from high to medium. On July 21, Gemini 3.6 Flash introduced changes in response to feedback about verbosity. Both updates could change answers without producing a newly published cutoff date.

Cutoffs and deprecation

Gemini deprecation dates describe model support and availability, not training freshness. Google uses deprecation for the end of support and shutdown for the point when an endpoint becomes unavailable.

Restricted access is a separate state. Google restricted Gemini 2.5 access for new users on Sept. 18, 2026, while explicitly saying the models were not deprecated and would continue to be served.

Stable also means suitable for production and usually unchanged. It does not mean permanent. Preview models can be deprecated with at least two weeks’ notice.

gemini-3-pro-preview shows the difference. The model had a documented January 2025 cutoff, then shut down on March 9, 2026, and was replaced by gemini-3.1-pro-preview. Its cutoff, release, replacement, and shutdown dates each answer a different question.

Monitor live Gemini visibility

A Gemini cutoff table cannot show whether current answers retrieve, cite, recommend, rank, or omit your brand. Live visibility depends on the prompt, market, intent, retrieval results, and model behavior.

In research covering July 1 to Aug. 31, 2025, Yext analyzed 6.8 million citations from 1.6 million questions per model. First-party websites supplied 52.15% of Gemini citations. Local websites contributed roughly 8% to nearly 20% across some query groups, reaching nearly 20% for unbranded, objective Gemini queries.

That local variation matters. A single global prompt set can hide what buyers see in London, Chicago, or São Paulo. Segment prompts by market and intent, then repeat them on a fixed cadence.

AI visibility measures how consistently a brand appears across sampled AI answers. Temso captures Gemini through live browser sessions, records full rendered responses and cited URLs, and repeats prompts while controlling location, language, session state, and cadence. You can monitor daily visibility, inspect citation sources, benchmark competitors, and prioritize gaps that need action.

Treat visibility scores as trends in sampled answers rather than exact daily readings or audience-reach figures. Start monitoring Gemini visibility with Temso’s AI search analytics, then compare the same market and intent segments week after week.