Claude と GPT の知識カットオフと学習期間を調査する手法

We can learn hidden facts about how frontier models were trained by “probing” them with carefully curated requests.

• By scoring them on niche facts we can approximate how many parameters models like GPT-5 and Opus have, using “Incompressible Knowledge Probes”
• By measuring how the models break down tokens we can reveal facts about the datasets mixtures they used to train the model (or at least the tokenizer) using “Data Mixture Inference”
• By scoring them on date or self-identification related questions you can also estimate training timelines (this post)

Everything here is an estimate. It’s possible that some speculation in this post is totally incorrect given there’s not a ton of publicly available ground truth to verify against.

As a brief primer (see Alex Wa’s blog for more), how we train massive large language models has converaged into 3 stages:

• Take a massive amount of general purpose data (aka scrape the internet) and “pre-train” a massive auto-complete model on that data.
• Use domain-specific “textbook quality” data to improve the base models and extend certain base capabilities like long-text understanding
• Turn the base model into the “assistant” persona, honing in on its personality, reasoning ability, and tool-calling.

While increasingly more compute is spent on post-training for boosting a model’s reasoning and problem solving, one of the most expensive and data-intensive steps is generating that pre-training checkpoint (by ‘checkpoint’ think of a massive claude-super-secret-2026-11-01-base.cpkt file).

While all labs operate slightly differently, what you might see (~page 44) is:

• The “pre-training” team kicks off and babysits a multi-month run to get a base checkpoint. These pre-training models often, but not always, imply major versions of released models (GPT-4 → GPT-5).
• While that’s happening, the “capability” and “post-training” teams will run experiments for how to improve on the most recent base model. Advancements in post-training and capabilities often manifest as minor versions of released models. These teams often also “distill” a single post-trained model into smaller variants that become model families (Fable/Opus/Sonnet/Haiku, Sol/Terra/Luna). Labs may also release post-trained models from half-baked pre-training checkpoints as soon as x% of the version N+1 checkpoint is better than the 100% baked version N checkpoint.
• The model released to the public is the culmination of the most recent checkpoint with the best set of capabilities and post-training techniques applied to it.

With this in mind, I was curious how much of this process you can “see” just by probing the model over the official APIs.

To estimate the pre-training checkpoint dates, I constructed a dataset of daily-facts from Wikipedia (e.g. 2025 in the United States) and gave every model an 8-way multiple choice quiz on what happened on a given day. Then, by analyzing the error rate timeline, you can see roughly when it loses signal from its training dataset.

You can then plot this for all models.

After staring at these charts for a bit, here’s what I’m speculating:

• Anthropic models Opus 4.7 onwards are all from the same training run that cuts off just around late December 2025. This is derived from how they all share a very similar effective knowledge cutoff (green). A core assumption I’m making here is that the pre-training base model completion date is highly correlated with the dataset timespan used, if that’s wrong these results could be off by some offset (e.g. it’s actually Jan 2026). It’s also interesting that Opus 4.7+ models all have a published reliable and overall knowledge cutoff that’s identical — maybe that’s due to a new post-training technique being used?
• OpenAI’s GPT-5.6 family comes from their own checkpoint, separate from GPT-5.5, that finished around late February 2026. This is derived from how they have a distinct effective knowledge cutoff from previous models. You’ll notice Luna looks like it can predict the future — that’s more of an artifact of it having a high error rate all around at a “low” reasoning effort.
• Opus 5 is a bit unusual. The published reliable and overall knowledge cutoffs are May 2026 and yet it seems to know nothing more than previous Jan 2026 cutoff models. I did several ablations to test whether it was an artifact of the types of probing questions I used, but not really — the cutoff applies to recall on coding package versions as well.

What if you just ask the model what today is?

It ends up being fairly correlated with fact-based estimates. If you look closely you can see some vertical lines within a few of the families of models.

• GPT-4.1 nano → GPT-4.1 mini → GPT-4.1
• Opus 4.7 → Sonnet 5 → Fable/Opus 5

Interpreting this graph as X = “pre-training corpus” and Y = “post-trained behavior”, these vertical strips (X constant, Y increasing) visualize active post-training on recency-biased datasets. Potentially distillation from old copies of teacher models is what causes smaller models to self-report older dates.

You can also make predictions on training timelines and datasets indirectly by looking at who the models think they are. The more a model sees “I am X” in its pre-training dataset, the more likely it is to repeat that when pushed and given no other grounding context.

After staring at these charts for a bit:

• Vertical bands show clear patterns of labs training on past-model outputs (from users). For OpenAI it’s GPT-4, GPT-4o, GPT-4.1 for a bit, GPT-5 and “ChatGPT” most recently. For Anthropic it’s 3.5 Sonnet then more recent models swap to Sonnet 4.5. This seems to align pretty well with training on chats from ChatGPT.com and Claude.ai respectively, where users chatted with the latest model and whose sessions became training material (directly or via web contamination). It seems unlikely to me these are coming from internal synthetic datasets given those would be much easier to suppress model identity (vs being embedded in the system prompt in the consumer chat sessions). Training-on-chats isn’t novel information but it is interesting to see expressed literally with probing like this.
• It’s a bit spicy that OpenAI models never identify as another lab’s model (besides briefly a Tesla Model S) yet Anthropic’s Sonnet 5 will regularly self-identify as GPT-4. It feels very unlikely that they are intentionally distilling GPT-4 but it’s possible a bunch of older ChatGPT chats are still making their way into the Claude training mixtures. It could also be carried through the generations via Sonnet 3.5 lineage (i.e. Sonnet 5 is trained on Sonnet 3.5 data which itself might have had a very GPT-4 heavy dataset). On top of this, in a follow-up experiment, when asked to answer identity questions “as model X would,” Claudes reproduce OpenAI models’ measured quirks at 68%; OpenAI models manage 8% on Claudes.

That’s it. Hope that was mildly interesting and if you want to explore the data a bit, here are the links:

• Model knowledge timeline
• What month does each model think it is?
• Does the model know its own name?
• Identity vintage: which era does each model think it is from?