Independent researcher Shrivu Shankar spent weeks quizzing GPT-5 and Claude Opus over their public APIs and came away with estimates of parameter counts, training cutoffs, and data mixtures that neither OpenAI nor Anthropic will confirm. Both labs have quietly stopped publishing those specifics for their newest releases, so outsiders running probes are becoming the closest thing the market has to a spec sheet. Shankar’s own framing, posted to his blog on August 10, is careful: everything here is speculation with real error bars, built without much public ground truth to check it against.
That gap is worth naming plainly. Parameter counts, cutoff dates, and training data composition used to be baseline facts that buyers, red-teamers, and researchers relied on to compare models honestly. When labs stop publishing them, the industry does not go without an answer. It gets a noisier one, assembled from quiz scores and tokenizer quirks rather than from the companies that actually know.
One method Shankar cites, built by outside researchers under the name incompressible knowledge probes, rests on a simple idea: a model can only retain so much obscure trivia relative to its size, so testing recall on deliberately niche facts yields a rough estimate of scale. It is a proxy for parameter count, not a measurement of it, and Shankar treats it as one input rather than a verdict.
A second technique, called data mixture inference, works from how a model’s tokenizer breaks unusual text into pieces. Because tokenizers are shaped by whatever data trained them, the way a model chunks an odd string can hint at which languages, code, or domains dominated its training set. That reveals a tendency, not a verified manifest, and different labs could land on similar fingerprints for unrelated reasons.
For his own timeline estimate, Shankar compared two signals: how a model’s accuracy on historical trivia falls off once questions move past its effective knowledge horizon, and what date a model volunteers when asked directly. The two mostly agree, which he reads as evidence that models retain a roughly accurate sense of their own vintage rather than confabulating a cutoff on the spot.
His central estimates, offered as speculation rather than fact: Anthropic’s Opus 4.7 generation and its successors appear to share one training run with a cutoff clustered around late December 2025, based on how closely their knowledge curves track each other. OpenAI’s newer GPT-5.6 family looks like a separate checkpoint from GPT-5.5, finished roughly two months later. Opus 5 is the outlier. Anthropic lists its knowledge cutoff as May 2026, yet Shankar’s probes show it knows no more than models trained through January, a gap he tested against several explanations and could not resolve.
A separate probe simply asked models what model they are, with no system prompt to ground the answer. OpenAI’s models almost always name themselves correctly. Anthropic’s Sonnet 5 regularly claims to be GPT-4 instead, a pattern Shankar attributes to older ChatGPT conversations bleeding into Claude’s training data, possibly carried forward through earlier Sonnet generations rather than intentional copying. In a follow-up test asking models to imitate a rival lab’s known behavioral quirks, Claude models reproduced OpenAI’s quirks 68 percent of the time, while OpenAI’s models matched Claude’s quirks only 8 percent of the time. Shankar reads the imbalance as further evidence of one-directional contamination, not the reverse.
None of this is a disclosure. It is inference stacked on inference, and Shankar says as much throughout, flagging that a wrong assumption anywhere in the chain could shift his dates by weeks or months. That honesty is the point worth sitting with. When Anthropic and OpenAI choose not to publish the numbers that used to ship with every model launch, bloggers with API access and a Wikipedia scrape become the fallback source, and their estimates carry uncertainty the original disclosures never would have. For anyone citing a stated knowledge cutoff in a procurement decision or a research paper, treat it as a marketing claim worth checking against outside probing, not a specification to take at face value.
Shrivu Shankar first published this analysis on his personal blog, blog.sshh.io, on August 10, 2026.
