Tokenization
BPE's long reign, the vocabulary-size race, and the byte-level challengers that have not yet taken the frontier.
Tokenization is the stack’s most stable layer. The algorithm GPT-2 shipped in 2019, byte-level BPE, is still what frontier models use in 2026. What changed is the vocabulary size, the training corpus behind it, and a growing research case that the layer should not exist at all.
The lineage
Figure 1. One algorithm, growing vocabularies. The byte-level challengers attack the layer’s existence rather than its parameters, and none has shipped in a frontier model yet.
The vocabulary race
| Tokenizer generation | Years | Vocab size | Note |
|---|---|---|---|
| GPT-2 / GPT-3 BPE | 2019 to 2020 | ~50K | English-heavy merges |
| Llama 1/2 SentencePiece | 2023 | 32K | Small vocab, high fertility on non-English |
| cl100k (GPT-4) | 2023 | ~100K | Code-aware |
| Llama 3 | 2024 | 128K | Switched to tiktoken-style BPE |
| o200k (GPT-4o) | 2024 | ~200K | Multilingual rebalance |
| Gemma / Gemini lineage | 2024 to 2026 | 256K | The current upper end |
Larger vocabularies buy compression (fewer tokens per document, so more effective context and cheaper serving) at the cost of embedding-table parameters and rare-token undertraining.
Persistent failure modes
| Failure mode | What happens |
|---|---|
| Fertility inequality | Languages distant from the merge corpus pay 2x to 4x more tokens for the same content, silently pricing out non-English users per character of output |
| Numeric splitting | Inconsistent digit tokenization degrades arithmetic; modern tokenizers force single-digit or fixed-width splits |
| Glitch tokens | Undertrained vocabulary entries produce anomalous behavior; documented across GPT-2 through Llama lineages |
| Boundary artifacts | Trailing-whitespace and prefix-space asymmetries remain a source of subtle prompt bugs in every BPE model |
Ledger
| Verdict | Techniques |
|---|---|
| Dead | Word-level vocabularies; character-level modeling at scale; 32K vocabularies at the frontier |
| Current | Byte-fallback BPE at 100K to 256K, digit-aware splitting, multilingual-weighted merge corpora |
| Contested | Byte-latent / patch-based models (research-strong, unshipped); “tokenization causes X” claims pending byte-level frontier evidence |
Next: Training Data, the corpus those tokens come from.