OpenAI's Ted Sanders explains model naming: version numbers encode release order and improvement size
Simon Willison · rss · 2026-10-29
Simon Willison quotes Ted Sanders of OpenAI explaining the company's model naming conventions.
Sanders notes that LLM performance is highly multidimensional—a model best at math isn't necessarily best at jokes, so no fixed ordering exists. Still, much of the variation is explained by a single variable: model size. Bigger models generally (a) perform more reliably, (b) cost more per token, and (c) recall esoteric facts better. OpenAI's models are ordered internally as Luna < Terra < Sol < Astra.
On version numbers, he says a version now signifies three things at once: release order, relationship to other releases, and magnitude of improvement—e.g. 6.1 launched after 5.6, may share training data/personality with other 6.x models (hence not 5.7), and isn't a big enough leap to merit a 7.
More from Models
- KV cache per token shrank 437x in 3 years — why this dev thinks AI gets cheaper — gnukeith · 2026-10-29
- DeepSeek V4.1 Flash runs 1M context in under 0.9GB RAM vs 320GB for Llama 70B — gnukeith · 2026-10-29
- Leaked Google docs page shows Gemini 4 Argon, fueling Gemini 4 Carbon rumors — nlarusstone · 2026-10-29
- 5 frontier LLMs grade 39 local models on 50 prompts; Bonsai-2 27B PQ2 tops the table — yobarisushcatel · 2026-10-29
- Report: Moonshot distilled OpenAI's Astra to train Kimi K3.1 Pro — kimmonismus · 2026-10-29
- Across 28 models, knowledge breadth tracks intelligence at Pearson r=0.86 — teortaxesTex · 2026-10-29