AI training data

What is a training data cutoff?

A training data cutoff is where a model's knowledge stops; here is how far behind models ship and what that means for data sellers.

By

Published · 6 min read

Illustration: a clipper ship under full sail against a low moon

The short answer

A training data cutoff is the last date covered by the data a model was trained on. Anything later is missing unless a search tool or a newer model supplies it. Anthropic's recent models shipped two to five months after their stated cutoffs, and research shows models do worse on text written after training, so recent, dated data matters to AI buyers.

What is a training data cutoff?

A training data cutoff is the date where a model's training data ends. A model learns from a fixed collection of text, code and other material, then its weights are frozen. Anything written later reaches it only through a search tool or a newer model.

The path from a model's training data to its release, as described on this page.

Training data cutoff vs knowledge cutoff

Some providers publish two dates. Anthropic's models overview lists a reliable knowledge cutoff, the date through which a model's knowledge is most extensive and reliable, and a training data cutoff, the broader range of data used. For its four current models, both read June 2026.

How far behind is a model on the day it ships?

Two to five months, for the eight Claude models we could match: we set each stated knowledge cutoff against the release month Anthropic's Transparency Hub gives. No glossary we read measures this gap.

Compiled by us from Anthropic's models overview (5.5 cutoffs) and Transparency Hub (release months, earlier cutoffs), both checked 2026-11-14. Gaps count whole months.
ModelStated knowledge cutoffReleasedGap
Claude Opus 4.7January 2026April 20263 months
Claude Opus 4.8January 2026May 20264 months
Claude Sonnet 5January 2026June 20265 months
Claude Fable 5January 2026June 20265 months
Claude Opus 5May 2026July 20262 months
Claude Opus 5.5June 2026September 20263 months
Claude Sonnet 5.5June 2026September 20263 months
Claude Haiku 5.5June 2026October 20264 months
Months from stated knowledge cutoff to release, Claude models
Months from stated knowledge cutoff to release, Claude models
Opus 4.73 months
Opus 4.84 months
Sonnet 55 months
Fable 55 months
Opus 52 months
Opus 5.53 months
Sonnet 5.53 months
Haiku 5.54 months

Source: Anthropic models overview and Transparency Hub, checked 2026-11-14. Compiled by DataBroker.

Four releases between April and June 2026 shared one January 2026 cutoff, so a newer model is not always a fresher one. And a model keeps its cutoff while it is offered: Claude Opus 5.5 will not be retired before September 22, 2027, when its June 2026 knowledge will be over a year old.

From Anthropic's models overview and Transparency Hub, checked 2026-11-14.

Is the stated cutoff the real one?

Not always. A 2024 paper by Jeffrey Cheng and colleagues, "Dated Data", defines an effective cutoff for each resource and topic a model learned, and finds it often differs from the cutoff the developer reports.

The authors trace two causes in open pretraining datasets: newer web crawl dumps carry a lot of old pages, and deduplication misses near-duplicates, so older versions of a page can outnumber the current one.

The two causes the Dated Data paper traces, drawn by us.

Why does a cutoff make fresh data worth more?

Because models get worse at the world the further it moves past their training data, and new data is the main fix. That is the finding of "Mind the Gap", a paper by Angeliki Lazaridou and colleagues presented at NeurIPS 2021.

The authors tested Transformer-XL models on text written after their training period. Performance dropped, and the drop grew with time. Bigger models did not solve it. Models that kept updating their knowledge with new information did reduce the decline.

Findings of Lazaridou et al., NeurIPS 2021, summarised by us.

What buyers have said about freshness

When Google expanded its partnership with Reddit on February 22, 2024, its announcement said Reddit's Data API delivers real-time, structured, unique content and gives Google structured access to fresher information to display, train on and otherwise use. It bought a stream, not a one-off archive.

Labs also use more than the public web. Anthropic's Transparency Hub describes recent Claude models' training data as publicly available information plus public and private datasets, user data where permitted, and synthetic data generated by other models.

What should a data seller do about cutoffs?

Treat dates as part of the product.

  1. 1

    Keep a date on every record

    Creation and last-edit timestamps let a buyer filter to what came after a model's cutoff. Don't strip them during cleaning.

  2. 2

    Show the date range in the sample

    State the earliest and latest record, and how many fall in each year, in the dataset card or cover note.

  3. 3

    Split the archive from the feed

    History is a one-off sale. New records as they are created can be a recurring licence, the shape of the Reddit and Google deal.

  4. 4

    Check rights to future data

    Consent and contracts that cover past records may not cover records collected next year. Confirm both before offering a feed.

  5. 5

    Start from published prices

    Our AI training data value estimator multiplies your amounts by published prices AI buyers paid, with each source cited.

AI training data value is what an AI company would pay to license your data for training models. Few prices are public; the reported ones are per minute of video, per photo, per hour of audio and per book. Our AI data licensing guide covers how one-off and recurring licences are structured.

What this answer doesn't cover

Our table covers one developer's models, using the months it publishes; we did not compile other labs. No published price we have found shows a premium for data after a given cutoff, so this post argues from research and from what buyers have said, not from prices.

Whether your data can be licensed at all depends on how it was collected and the promises made to the people in it.

Sources

5 sources · checked Nov 15, 2026

The DataBroker team checked the facts on this page against 5 public sources, last on .

  1. Anthropic, models overviewReliable knowledge and training data cutoffs (June 2026) for the 5.5 models and their definitions; Opus 5.5 retirement date · Checked
  2. Anthropic Transparency HubRelease months; knowledge cutoffs for Opus 4.7, Opus 4.8, Sonnet 5, Fable 5 and Opus 5; training data descriptions · Checked
  3. Cheng et al., "Dated Data: Tracing Knowledge Cutoffs in Large Language Models" (2024)Effective versus reported cutoffs; old data in new crawl dumps; deduplication gaps · Checked
  4. Lazaridou et al., "Mind the Gap: Assessing Temporal Generalization in Neural Language Models" (NeurIPS 2021)Worse performance on later text; scale doesn't fix it; updating with new data helps · Checked
  5. Google, "An expanded partnership with Reddit", February 22, 2024Real-time Data API access and fresher information for training and display · Checked

What is a training data cutoff: FAQ

Can a model learn about events after its cutoff?

Not by itself. A search tool, retrieval or a document pasted into the conversation can give it newer information for that answer, but its own knowledge only changes when it is trained again.

Why is a model released months after its cutoff?

Training, testing and safety work take time after the data is collected. For eight recent Claude models the gap between stated knowledge cutoff and release was two to five months.

Does a model know its own cutoff?

Don't rely on it. Researchers have found that a model's effective cutoff often differs from the one its developer states, so check the developer's documentation rather than asking the model.

Is old data worthless to AI labs once it's in a model?

No. Data that was never public was never in a crawl, whatever its age, and labs say they train on private datasets. A cutoff matters most for public material, which a model may already have seen.

Related

  • AI training data

    AI training data licensing explained: which datasets labs want, publicly reported deal values, and the copyright and transparency rules that now apply.

  • AI training data value estimator

    Estimate AI training data value from published prices only: choose what your company holds and see what AI buyers reportedly paid for the same kind.

  • LLM training data

    LLM training data comes in three kinds, for three stages of building a language model, and each raises different rights questions for sellers.

More from the blog

Free tool

Try the AI training data value estimator

AI training data value is what an AI company would pay to license your data for training models. Few prices are public; the reported ones are per minute of video, per photo, per hour of audio and per book.

  • Free
  • No sign-up
  • Runs in your browser