Editing Openai/69408af5-1bf8-800f-b5dc-86cd01e07ec0 (section)

==== ### ====

Estimated: ~15–30+ trillion tokens
* Gemma (a smaller, open sibling) is documented at ~11T
* Gemini is multimodal, closed, and Google’s flagship
* Almost certainly trained on more data than Gemma
* Likely the largest overall training corpus in existence

===== Estimated: ~15 trillion tokens =====
* Meta publicly stated “15T+ tokens” for LLaMA 3.1
* This is one of the few frontier-scale figures Meta has openly acknowledged
* Mostly text and code, less multimodal than Gemini

===== Estimated: ~10–20+ trillion tokens =====
* OpenAI does not publish token counts
* Multiple credible estimates place GPT-4.x in low-to-mid double-digit trillions
* Includes heavy synthetic data, code, multimodal inputs
* Could plausibly rival Gemini but lacks confirmation

===== Documented: ~11 trillion tokens =====
* Explicitly stated in Google model cards
* This is a hard lower bound reference point
* Not multimodal at Gemini’s scale

===== Estimated: ~5–10 trillion tokens =====
* Anthropic is intentionally opaque
* Claude’s performance suggests multi-trillion scale
* Likely smaller than GPT-4/Gemini but still frontier-class

===== Estimated: ~5–8 trillion tokens =====
* xAI has publicly emphasized scale and rapid iteration
* Training data likely includes large proprietary and real-time sources
* Less historical depth than Google or OpenAI

===== Documented: ~3.5 trillion tokens =====
* Largest publicly audited training dataset
* Pure text focus
* Still the benchmark for “confirmed” scale

===== Estimated: ~2–4 trillion tokens =====
* Mistral has not disclosed totals
* Performance suggests multi-trillion training
* Smaller compute budgets than Big Tech labs

===== Estimated: ~2–3 trillion tokens =====
* Amazon does not disclose numbers
* Training scale inferred from enterprise deployment and model size
* Likely conservative compared to Gemini or GPT

===== Estimated: ~3–4 trillion tokens =====
* Older than Gemini
* Well-documented large-scale training
* Superseded but still massive