Source time Edition date
IndustrySimon Willison
Quoting Ted Sanders
引用 Ted Sanders
Ted Sanders via Simon Willison is worth a quick look because the value of the quote hinges on context Simon supplies rather than the words alone, so readers must decide how much weight to give it.
Judgment
AI product and platform teams choosing an OpenAI model tier should treat Luna < Terra < Sol < Astra as an ordinal size/cost ladder, not a guarantee that the larger model wins every task; the tradeoff is higher per-token cost for reliability and esoteric recall.
Basis
Ted Sanders, quoted by Simon Willison on 28th October 2026, says LLM performance is multidimensional: there is no fixed ordering where the math-best model is also the jokes-best model. He says much performance variation is explained by one variable, model size: bigger models are generally more performant/reliable, cost more per token, and have better recall of esoteric facts. He lists OpenAI's order as Luna < Terra < Sol < Astra. Version numbers, he says, encode release order, relationship to other releases, and magnitude of improvement: 6.1 came after 5.6, may share training data/personality with other 6.1 models (so not 5.7), and is not a big enough leap to merit a 7. That makes the naming a signal about tier and cost, not a task-by-task capability ranking.
Unclear The source does not give benchmark scores, per-token prices, or latency figures for Luna, Terra, Sol, or Astra, so the cost-performance tradeoff cannot be quantified from the quote alone.
Next Check OpenAI's public model documentation for per-token pricing and task-specific benchmark results for Luna, Terra, Sol, and Astra before treating the size ordering as a procurement rule.