Every AI has a freshness date.
Here's the board.

This page is a board. One row for each big-name AI chat tool, showing the date its built-in knowledge stops. Bookmark it, check it before you trust anything an AI tells you that has a date attached, and come back — we re-check it and re-stamp it.

First, the term. Every AI chat tool has a knowledge cutoff: the date past which you can't count on its built-in knowledge, set when the company finished training it. Ask about anything after that date and you can't trust it to know — and it won't say "I don't know." It guesses, smoothly, in the same confident voice it uses for everything else.

That's the whole reason this board exists. The tools don't put their expiration date on the label. The vendors publish it — some of them — buried in developer documentation you were never meant to read. We read it. Here it is.

The board

Every row below was read live off the vendor's own pages on July 20, 2026. "Everyday model" means the engine behind the standard chat window most people get — the model glossary is on the how-AI-works page.

Tool	Company	Everyday model	Knowledge cutoff	What that means for you
ChatGPT	OpenAI	GPT-5.5 Instant — OpenAI's own release notes call it the default	December 1, 2025 — the date OpenAI publishes for GPT-5.5, the engine behind the Instant default. (OpenAI doesn't print a separate date for the Instant version itself; this is the closest one on record.)	Anything from 2026 — a price, a product, a code change — it's guessing.
Claude	Anthropic	Sonnet 5 — Anthropic's June 30, 2026 announcement names it the default for Free and Pro	January 2026	Months old the day you read this, same as every row here.
Gemini	Google	Gemini 3.5 Flash — named the app default in Google's May 19, 2026 announcement	January 2025 — and the same Google page also says "Latest update: May 2026." Google prints both; it doesn't say which one covers your question.	Treat January 2025 as where its solid knowledge ends. Don't bet a job on the update covering what you need.
Copilot	Microsoft	Not published — Microsoft's Copilot pages name no model, as of the July 20, 2026 check	Not published	No date means no promise. Treat it as stale.
Grok	xAI	Grok 4.5 — xAI's docs call it their flagship	February 1, 2026 — xAI's docs state it verbatim	Same rule as the rest: February forward, it's guessing.
Meta AI	Meta	Not published — Meta's docs list the model's specs with no cutoff field at all	Not published	No date means no promise. Treat it as stale.

One credit where it's due: xAI's own documentation says plainly that "Grok has no knowledge of current events or data beyond what was present in its training data." That's the straightest sentence any vendor on this board has written about its own product. It applies to every row.

The rows that say "not published"

Two vendors on this board don't tell you when their tool's knowledge stops. That's not a gap in our research — the absence is the finding. Microsoft and Meta, as of the July 20, 2026 check, publish no cutoff date for their everyday chat tools.

You already know what to do with a product that won't state its own spec. You don't assume the best number. You assume nothing, and you verify. Treat an unpublished cutoff as stale until proven otherwise.

The rule this board exists to teach

Anything with a date in it — prices, codes, rules, products, suppliers, phone numbers — you verify live before you use it. Look it up yourself, at the source, today. The cutoff is why. The tool's knowledge stopped months ago, it doesn't know what changed since, and it will not warn you. The AI is a helper with an old code book, and the checking never transfers.

The fine print on the fine print

Even the cutoff date has fine print. Anthropic's own model documentation, as of July 20, 2026, gives two dates per model and defines them like this: "Reliable knowledge cutoff indicates the date through which a model's knowledge is most extensive and reliable. Training data cutoff is the broader date range of training data used." In plain terms: one date is where the knowledge is solid, the other is just how far back the raw training text reached. The board above prints the first kind — the stricter one. (For Claude's January 2026, both dates happen to match.)

Read that carefully. A published cutoff marks the edge of the model's most reliable knowledge, not a wall where knowledge stops clean. Past that date, the vendor doesn't promise the knowledge holds up — so treat the date as where you stop trusting it. Which lands you back on the same rule: dated facts get verified live, every time.

That's the deal on this page. We re-check the board and re-stamp the date at the top of that box, and you keep the one rule that outlives every row on it: if it has a date in it, verify it live before you use it.

Next: how AI works — why the machine guesses instead of saying "I don't know."