Marcelo KunzeAsk an AI who runs a company, what its product costs, or how much it raised, and you get a fluent answer with a date missing. The answer comes from the model's memory, and the memory stopped months ago.
This isn't a bug you can wait out. It's how the models are built, and it has three fixes, each covering more than the last.
This post covers what a training cutoff is, why web search closes less of the gap than it seems, how to spot a stale answer, and the three fixes in order of effort.
What a training cutoff is, and how old your AI really is
A model is trained on a snapshot of text that ends on a date, the training cutoff. Then it's tested, tuned and released, which takes months. On launch day it is already behind, and it never learns another thing: every answer from memory is as of the cutoff, for the rest of the model's life.
The published numbers, as of November 2026: GPT 5.6 shipped in July 2026 with a cutoff of February 2026. Claude Fable 5 shipped in June 2026 with a reliable cutoff of January 2026. Five months behind on day one, and nine to ten months behind as you read this.
The worse part is that a model doesn't know what it missed. Ask about something after its cutoff and it won't say it doesn't know; it reasons from what it has and hands you a plausible answer with the same confidence as a fact. That's the mechanism behind an AI inventing a funding round or quoting last year's price.
A model never says it's out of date. It answers as if it isn't.
Why web search doesn't fix it
Search was supposed to close the gap, and for headlines it does. Three things limit it.
It only runs when the model decides. ChatGPT and Claude both choose whether a question needs a search. A question the model thinks it already knows gets answered from memory, cutoff and all, with no search at all.
It only reaches the free web. A search tool reads what a search engine indexed: free pages, the same ones every other AI reads. It can't log in, can't pay, and can't open the data that answers business questions: current prices across stores, who raised what this quarter, who holds which role, what people are saying in a thread four hundred comments deep.
It reads ten snippets. A search pass skims the top results and summarizes them. Where the answer is a table in a database, ten snippets don't contain it.
Every chatbot reads the same free web, so every chatbot gives the same answer, at the same time.
How to catch a stale answer
Three questions expose one.
- As of when? Ask for the date behind each fact. A model answering from memory either gives its cutoff or dodges.
- Where did that come from? Ask for a link behind every claim. No link means no source, which means no trust.
- What couldn't you verify? A good answer admits its gaps. A stale one has none.
You can build all three into one prompt, which is fix one.
Fix one: a prompt that forces a search and dates every claim
The cheapest fix, and it works for anything on the free web. Paste this ahead of a time-sensitive question, with today's date filled in:
Today is {date}. Before answering, search the web.
For every fact, give the source link and the date it was published.
If you can't find a current source, say so instead of answering from memory.
End with a list of what you couldn't verify.What it buys you: the model searches instead of guessing, and the dates show at a glance which facts are current. What it doesn't: anything behind a login or a paywall stays out of reach, and the model can still misread a page.
Fix two: give it the document
If the current data is yours (a report, an export, a CSV from a tool you pay for), hand it over. Upload it to the chat, paste it, or in Cursor drag it into context, and ask the question against the file. For developers this is retrieval-augmented generation, RAG: your documents are indexed, the relevant parts are pulled into the prompt, and the model answers from them.
It works, it's accurate, and it's manual: you export, you upload, you repeat tomorrow. It covers data you already have. It does nothing for data you don't.
Fix three: connect a live source it can query mid-answer
The structural fix. MCP (Model Context Protocol) is an open standard that lets AI apps use outside tools. Add an MCP server as a connector, and the model can call it while it writes: you ask, it notices it needs current data, fetches it, and answers from the result. No export, no re-upload, no cutoff.
There are three kinds of connector (a vendor's own server, your own API keys behind a custom server, and one connection that pays per call), and the setup for Claude, ChatGPT and Cursor is in the connector guide.
Frames is the pay-per-call kind, built for the data behind paywalls. Your AI searches a catalog of about 37,000 paid tools for free, picks one, and Frames pays the provider for that call, in stablecoins, then checks that the data arrived. You pay the cost of the data plus 15% in credits, with one balance and no API keys.
- You ask your AI
- Your AI searches Frames, free
- Frames pays the provider
- Frames checks the data arrived
- Your AI gets the data and a receipt
Every paid call is checked. Empty results, errors dressed up as success and calls that fail the delivery check aren't charged, and each request runs under a cap your plan sets. Each purchase gets a receipt: the tool, the price, the payment network and transaction, and whether it delivered. In Claude Code the whole setup is one line:
claude mcp add --transport http --scope user frames https://api.frames.ag/mcpClaude, ChatGPT, Cursor, Codex and the rest are on frames.ag/docs/mcp.
Which fix a question needs
The fix follows from where the answer lives.
| The question | Where the answer lives | The fix |
|---|---|---|
| What happened in a market this week | The free web | One: search, with dates |
| What our Q3 numbers say | Your own files | Two: give it the document |
| What a competitor charges today | Paid retail data | Three: a live connector |
| Who raised a Series A this quarter | Funding databases | Three: a live connector |
| What people on X say about our launch | Full social threads | Three: a live connector |
Whichever fix you use, keep the three questions from above. Even a connected answer deserves a date and a link behind every number.
What the fixes cost
Fixes one and two cost nothing but your time. Fix three depends on the connector: a vendor's own server costs that vendor's subscription, whether you query it or not.
Frames starts free. The Free plan gives you 3,000 credits a month with no card, and every plan reaches the whole catalog. 1 credit is $0.001, searching is free, and a request is capped at 1,000 credits on the Free plan. Paid plans start at $19 a month for 10,000 credits; the full table is on the pricing page.
Out of date is the default
No model will ever be current on its own; the training cutoff guarantees it. Search covers the free web, when the model bothers. For everything else, current is something you connect, not something you wait for, and it's one connector away.
If a fix doesn't work in your app, or a source you need isn't in the catalog, find us on Telegram.
It depends on the model. GPT 5.6, the current one as of November 2026, has a published cutoff of February 2026. Ask ChatGPT and it will tell you, roughly; the model's page has the exact date.
Anything after the cutoff reaches an answer only through search or a connector, never from memory.
It can search the free web, and it does so when it decides a question needs it. You can force a search with the search option or with a prompt like the one in fix one.
It has no access to paid data (prices across stores, funding databases, contact data, on-chain activity) unless you add a connector that reaches it.
Yes, with web search turned on, which Claude uses when it judges a question needs fresh information. Plenty of answers still come straight from its training data.
Connectors let it reach sources search can't, including paid ones. On the free plan you get one custom connector; Team and Enterprise plans have an owner add it first.
No. RAG (retrieval-augmented generation) pulls from documents you already have and indexed, which is fix two. A connector queries a live source at answer time, so it covers data you don't have yet.
Many setups use both: your own documents for what you know, a connector for what changes.
On Frames, searching the catalog is free, and a paid call costs the data plus 15%, usually cents. 1 credit is $0.001.
The Free plan gives you 3,000 credits a month with no card, every request is capped at 1,000 credits, and calls that don't deliver aren't charged, so a month of occasional questions fits inside the grant.
