# AI knowledge cutoffs and live search, read from vendors

> A brand's recent launch can be missing from an AI answer for two different reasons. What the vendors document, and how to tell which. Read the guide.

**Category:** Guide · **Published:** 2026-10-16 · **Canonical:** https://www.zumihq.com/blog/ai-knowledge-cutoffs-and-live-search

## Reading brief

- **Decision:** Whether a wrong or missing brand fact in an AI answer is a memory gap, which time fixes, or a search gap, which a crawlable page fixes.
- **Evidence:** Cutoff dates and search behavior as OpenAI, Anthropic, and Google state them on their own pages, read on 16 October 2026.
- **Action:** Test a brand fact that postdates the engine's cutoff, check whether the answer cites pages, and fix the page, the crawler access, or the off-site sources accordingly.


On 16 October 2026, OpenAI's own pages gave GPT-6 Astra a knowledge cutoff of April 30, 2026, and GPT-6 Luna one of May 18, 2026. The cheaper model's training reaches about two and a half weeks past the flagship's.

Those dates matter to any brand that launched, renamed, or repriced something this year. An engine can miss the change because it never learned it, or because it did not look. The two causes need different fixes.

This guide reads what each vendor states about its own cutoffs and search, then turns that into a way to tell a remembered answer from a searched one.

## Key takeaways

- Every cutoff below comes from the vendor's own page and carries its read date. Third-party cutoff tables go stale within a release cycle, so none is used here.
- OpenAI publishes a cutoff per model: April 30, 2026 for GPT-6 Astra and GPT-6.1 Sol, May 18, 2026 for GPT-6 Luna, and April 20, 2026 for the model used in ChatGPT. Anthropic lists June 2026 for all four current Claude models.
- Google's Gemini API model pages carry no cutoff. The Gemini 3.8 Flash model card gives March 2026, and adds that in some domains knowledge may stop at January 2025.
- A cutoff only describes what a model learned in training. Anthropic, Google, and OpenAI each document a separate search tool that reaches past it, and Google states that the model decides when to use it.
- A brand fact that postdates the cutoff can reach an answer only through search. A fact the answer gets wrong from before the cutoff is a memory problem that no page edit fixes quickly.

## What do the vendors say about their own cutoffs?

A knowledge cutoff is the date after which a model has no training data. Vendors do not define it identically, and they do not all publish it.

**OpenAI** lists a cutoff on each model's page. GPT-6 Astra and GPT-6.1 Sol read "Apr 30, 2026 knowledge cutoff," GPT-6 Luna reads "May 18, 2026 knowledge cutoff," and the page for the latest instant model used in ChatGPT reads "Apr 20, 2026 knowledge cutoff" ([GPT-6 Astra](https://developers.openai.com/api/docs/models/gpt-6-astra), [GPT-6 Luna](https://developers.openai.com/api/docs/models/gpt-6-luna), [Chat Latest](https://developers.openai.com/api/docs/models/chat-latest), read 16 October 2026). The Chat Latest page adds that "the underlying model snapshot will be regularly updated," so that date can move without a new model name.

**Anthropic** publishes two terms per model. Its [models overview](https://platform.claude.com/docs/en/about-claude/models/overview), read 16 October 2026, lists a reliable knowledge cutoff and a training data cutoff, both June 2026 for Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5. The reliable cutoff is "the date through which the model's knowledge is most extensive and reliable." The training data cutoff is the broader range of data used, so Anthropic documents two dates where others give one.

**Google** is the least direct. The [Gemini API models page](https://ai.google.dev/gemini-api/docs/models) and the per-model page for Gemini 3.8 Flash, both read 16 October 2026, list token limits and capabilities but no cutoff. The [Gemini 3.8 Flash model card](https://deepmind.google/models/model-cards/gemini-3-8-flash/) states that the cutoff "is March 2026," and then adds that "in others they may experience the model's knowledge is limited to January 2025 (in line with the Gemini 3 Model Family)." One model, two dates, both from Google.

Perplexity, Copilot, Grok, DeepSeek, AI Overviews, and AI Mode were not read for this guide. Where a vendor publishes no cutoff, the honest entry is "undisclosed."

## Why is a third-party cutoff table the wrong source?

Cutoff tables on other sites are easy to find and tidy to read. Few say which vendor page a row came from, and none can follow a vendor that changes a date.

Two cases from the vendors themselves show the problem. OpenAI says the snapshot behind its ChatGPT model page will be regularly updated, so that date can move under an unchanged name. Google gives two dates for one model, a split that a one-column table cannot hold.

A brand that needs the cutoff for a particular engine should read the vendor's page, note the date of the read, and expect the answer to change. Any figure that cannot be found there should be written down as undisclosed.

## Does search reach past the cutoff?

Each vendor documents a search tool that sits beside the model, not inside its training. Anthropic says its web search tool gives Claude "real-time web content, allowing it to answer questions with up-to-date information beyond its knowledge cutoff" ([web search tool](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool)). Google says grounding with Google Search lets Gemini "cite verifiable sources beyond its knowledge cutoff" ([grounding documentation](https://ai.google.dev/gemini-api/docs/google-search)).

The OpenAI model pages above list web search among the tools each model supports.

Google documents that the model decides when to use it. Its documentation says the model "analyzes the prompt and determines if a Google Search can improve the answer," and, if needed, generates one or more search queries on its own. Whether a given question triggers a search is not something a brand can assume.

How ChatGPT's search finds sources once it does run is covered in [what OpenAI documents about ChatGPT search](/blog/how-chatgpt-search-picks-sources). Perplexity's source behavior is covered in [how Perplexity cites sources](/blog/how-perplexity-cites-sources).

No public figure in these pages says what share of answers uses search. A brand should not plan around one.

## How can a remembered answer be told from a searched one?

An answer to a question about a brand comes from memory, from search, or from both. The two age differently, and they break differently.

The check takes four steps:

1. **Pick a fact that postdates the cutoff.** A product launch, price change, or new office from the months after the vendor's stated date works. If the answer states it correctly, search almost certainly ran, because training could not have supplied it.
2. **Pick a fact that predates the cutoff.** A founding year or long-standing product line works. If the answer gets that wrong, the cause is memory, since the engine had the chance to learn it.
3. **Look for cited pages.** A list of sources under the answer is the practical sign that a search step ran. An answer with no sources may still have searched, so absence is a hint, not proof.
4. **Repeat the question.** Search results can shift from run to run. A fact that appears in some runs and not others suggests search is involved.

None of this is a measurement of how often engines search. It is a way to read one tracked answer correctly before deciding what to fix.

## What gets fixed once the cause is known?

A search gap has a concrete owner. The page that states the fact has to be reachable by the engine's crawler, load without a login, and say the fact in plain text, which the [AEO optimization checklist](/blog/aeo-optimization-checklist) walks through crawler by crawler. A brand whose products show up in ChatGPT answers should also read which pages those answers cite, as in the [shopping guide](/blog/chatgpt-shopping-guide).

A memory gap has no quick fix. The model learned something wrong or learned nothing, and the correction arrives with a later training run. The practical move is to get the right fact onto sources the engine is likely to find and cite in the meantime, and to track whether it appears in answers over time.

Treating the two as one problem wastes effort. Editing a page does not change what a model remembered, and waiting for a new model does not fix a page that no crawler can read.

This guide did not measure how often tracked answers use search, did not read the cutoffs for every engine, and does not cover the consumer apps' default search settings. Those pages change by plan and by release, and each needs its own dated read.

Zumi tracks which engines mention a brand and which sources they cite across [the major AI engines](/platform/ai-search), which is the evidence this reading depends on. A brand that wants that view for its own questions can [book a demo](/book-demo).