Why models miscount letters in "strawberry": they see tokens, not letters
Euphoric_North_745 · reddit · 2026-09-28
A Reddit post explains the root of the classic "how many R's in strawberry" failure: the ChatGPT app shows your typed text, but the API actually receives token ID sequences (e.g., [5299, 1991, 460, 382, 101830, 30]).
The model reasons in token numbers, not human letters — there are no letters in its "brain" — and its numeric output is converted back to text by the API. A translator sits between you and the model, which is why counting letters trips it up. The post links to OpenAI's tokenizer tool for verification.
More from Models
- GPT-6 Astra system card: CoT monitor recall drops below 11%, latent reasoning kills monitorability — enginetown · 2026-09-28
- User pays $60 in Codex credits after lockout prompt, still blocked — and OpenAI can't explain where the credits went — BrilliantNo76 · 2026-09-28
- Qwen Quietly Ships decision-model-preview, a Jev Rival With No Open Weights — pneuny · 2026-09-28
- Claude Opus 5.5 compiles a dossier and leak timeline of OpenAI's 'o' always-on agent — shaunralston · 2026-09-28
- LiveNerf repo tracks Opus 5.5 daily benchmarks to detect if Anthropic nerfs the model — TheOnlyVibemaster · 2026-09-28
- Paid users report ChatGPT blocks VPN access, leaving users in China locked out — pizzababa21 · 2026-09-28