Researcher Probes Frontier AI Models to Reveal Hidden Training Timelines and Shared Checkpoints
Summary
- • A researcher used Wikipedia trivia quizzes on live model APIs to estimate when Claude and GPT models were pre-trained.
- • Findings suggest Anthropic's Opus 4.7+ model family all share a single pre-training run with a cutoff of approximately late December 2025.
- • The analysis implies popular Claude variants (Sonnet, Haiku, Fable) are post-training distillations of one base checkpoint, not independently trained models.
- • While methodology is reproducible via public APIs, the author cautions findings are speculative without public ground-truth from the labs.
Details
Wikipedia quiz methodology
Researcher feeds models daily Wikipedia trivia; where error rates spike reveals the effective end of pre-training data, estimating knowledge cutoffs.
Anthropic Opus 4.7+ shared checkpoint
All Claude models from Opus 4.7 onward appear to share one pre-training run, estimated to have a dataset cutoff around late December 2025.
Model families as distillations
Claude's Sonnet/Haiku/Fable/Opus variants are likely created by distilling a single post-trained base checkpoint, not from separate pre-training runs.
Incompressible Knowledge Probes
Technique for estimating model parameter counts by scoring on niche factual questions that cannot be 'compressed away' during training.
Modern LLM training pipeline
Pre-training (months-long run) → domain-specific capability tuning → post-training (persona, reasoning, tool-calling). Post-training produces named model variants.
Release cadence vs. training innovation gap
Multiple model releases may mask that labs are shipping different post-training configurations of the same underlying base checkpoint.
Source: Independent researcher analysis published on Hacker News. All findings are estimates; no public ground truth available from Anthropic or OpenAI.
What This Means
This independent analysis offers a rare window into how frontier AI labs structure their training pipelines — suggesting that major model families may share more pre-training DNA than their marketing implies. The finding that Anthropic's Opus 4.7+ models may all stem from a single December 2025 checkpoint highlights the growing gap between model release cadence and actual fundamental training investment. While the methodology is creative and reproducible using only public APIs, the author appropriately cautions that without lab-published ground truth these remain estimates — but the approach itself signals a growing community effort to hold AI labs more accountable for transparency about their training practices.