No single 2026 token ratio exists — the one fully quantified frontier case puts pre-training ahead of post-training by about 2,434:1

Asked (summary):

For frontier LLMs released in 2026, what is the token ratio of human versus synthetic training data, and the token ratio of pre-training versus post-training?

The scan collected 17 model–source rows on 2026-era frontier models — official lab pages, technical reports and model cards, plus secondary analyses — covering pre-training token counts, post-training token counts and synthetic-data disclosures. Here “human” means human-origin source material and “synthetic” means model-generated text or code. The honest answer: labs almost never disclose total post-training tokens or full token provenance, so no defensible industry-wide ratio can be computed — only model-specific examples and lower bounds, shown below.

What labs actually disclose, model by model

Each row is one model–source pair; each column one disclosure. Colour marks how complete the disclosure is — hover or tap a cell for the source text.

exact figure lower bound / partial mixture share qualitative only not disclosed

Step 3.5 Flash is the one clearly quantified 2026 case: 17.6T pre-training tokens vs 7.23B post-training tokens — about 2,434:1, i.e. 99.959% vs 0.041% of tokens. arxiv.org

Nemotron 3 Ultra reports ~20T pre-training tokens and phase mixtures with at least 47.4% (phase 1) and 43.3% (phase 2) in explicitly synthetic or SFT-style categories — mixture shares, not an exact whole-run synthetic fraction; no total post-training token count. research.nvidia.com

Ling / Ring 2.6 discloses ~9.6T pre-training tokens and ~70B synthetic web-code tokens — a synthetic lower bound of 0.73% — while leaving the full synthetic share unknown. arxiv.org

MAI-Thinking-1 is described as 30T pre-training tokens with 100% human-origin pre-training data; its 3.55T figure is mid-training, not strict post-training, and SFT/RL token totals stay undisclosed. felloai.com

Inkling reports 45T pre-training tokens and only qualitative statements — synthetic SFT data plus RL on synthetic and human-created environments — with no token shares at all. thinkingmachines.ai

What this means in practice

State the answer as ranges and examples, not an average. Human:synthetic pre-training spans from a disclosed 100:0 human-only policy (MAI-Thinking-1, per secondary coverage) to roughly balanced mixtures in models that lean heavily on synthetic data (Nemotron 3 Ultra’s ≥43–47% synthetic/SFT-style phase shares), with most labs publishing only partial lower bounds or nothing. For strict pre-training:post-training, the one quantified frontier example is ~2,434:1; the industry-wide ratio is undisclosed. Post-training can still be compute-intensive despite tiny token counts, because RL generates many rollouts and carries inference and reward-model cost. Web data increasingly contains prior AI output, so even “human” web tokens have blurred provenance.

All 17 collected rows

Primary rows carry the substantive comparison; secondary rows are context only. The 2025 Ling 2.0 row is excluded from 2026 conclusions.

Model	Tier	Released	Pre-training tokens	Post-training tokens	Synthetic disclosure	Source

Method: 17 model–source rows assembled from a 2026 web scan of training-data disclosures (official lab pages, technical reports, model cards, arXiv papers and secondary analyses), one source URL per row. Token figures are as stated by each source; blanks mean not disclosed, never zero. The grid encodes disclosure completeness, not magnitude. This is a scan of found disclosures, not a census of all proprietary frontier models — major closed labs typically publish no token counts at all. MiMo-V2.6’s parsed pre-training number is malformed in the source extract, so only its source text is shown. Long disclosure notes were trimmed for space; follow each source link for detail.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.