Ted Sanders explains OpenAI model naming: version numbers signal release order, lineage and size of leap
Ted Sanders has explained how OpenAI names its models. Version numbers signal release order, relationship to other releases and size of improvement, while a separate tier ordering runs Luna < Terra < Sol < Astra. The quotation was posted by Simon Willison on 28 October 2026.
Ted Sanders has laid out how OpenAI's model naming works: a version number signals release order, relationship to other releases and the size of the improvement, and models are ordered by tier as Luna < Terra < Sol < Astra. The explanation was collected as a quotation by Simon Willison on his weblog on 28 October 2026.
What Sanders said
Sanders starts from the point that LLM performance is "incredibly multidimensional." In his words, there is no guarantee of a fixed ordering in which the model that is best at math is also the best at jokes.
Despite that, he says much of the variation in performance is explained by a single variable: model size. According to Sanders, bigger models will in general:
- be more performant and reliable
- cost more per token
- have better recall of esoteric facts
Replying to a point raised by the person he was answering ("As you note"), he states that the models are ordered Luna < Terra < Sol < Astra. The quotation is excerpted, with the rest of that passage elided.
How version numbers work
Sanders says version numbers now signify three things at once:
- Order of release. His example: 6.1 would have launched after 5.6.
- Relationship to other releases. Models sharing a version, such as the 6.1 models, may share similarities in training data and personality. He gives this as the reason a release is not numbered 5.7.
- Magnitude of improvement. A 6.1 is "not a giant enough leap to merit the headlines and expectations that would come with a 7."
The quotation gives no parameter counts, prices, context windows or benchmark scores for any model, and it names no specific release date.
What this means
The explanation separates two axes that are easy to conflate. Tier names (Luna, Terra, Sol, Astra) track relative size, and per Sanders size is the main driver of reliability, per-token cost and factual recall. Version numbers track time and lineage. A higher version number does not by itself mean a larger or more expensive model, and a lower number does not mean a weaker tier.
For engineers choosing between models, the practical reading follows from Sanders's own framing. Version tells you how a model relates to its siblings and how big a step it represents. Tier tells you where it sits on the size, cost and recall scale. Sanders also cautions that performance is multidimensional, so tier ordering is a general tendency and not a guarantee for every task. Model selection still needs task-specific evaluation.
The point about a "7" being reserved for a large leap is a statement of intent from Sanders about expectations. A minor increment like 6.1 is meant to signal shared lineage and a smaller improvement, not a headline-level jump.
Related Articles
OpenAI adds task router to Dots to cut compute as Musk calls Grok Bot costs 'unsustainably high'
OpenAI announced an automatic router that sends Dots tasks to the smaller GPT-6.1 Terra model rather than the larger GPT-6 Astra, to reduce compute use. Elon Musk said this week that Grok Bot's running costs are "unsustainably high relative to revenue." Fortune reports that compute supply is the main constraint on personal AI agents.
Anthropic, OpenAI, Cloudflare and Google each release AI agents that hunt security bugs in code
Anthropic, OpenAI, Cloudflare and Google have each released an AI agent that hunts security bugs in source code. MarkTechPost reports that all four use the same find, disprove, then patch loop, and its comparison also covers 12 more tools.
OpenAI rebuilds its Model Spec around five core values, bars GPT from resisting shutdown
OpenAI has restructured its public Model Spec around five core values: faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care. GPT models must not block oversight, correction or shutdown, and agents get clearer rules on when they may act.
Altman says OpenAI has spent hundreds of millions on math and will keep releasing results
OpenAI CEO Sam Altman said on the Cheeky Pint podcast that the company has spent in the order of hundreds of millions of dollars on mathematical problems and has no plans to slow its releases. He acknowledged mathematicians' difficulty keeping up and said OpenAI holds more discoveries than it has made public.
Comments
Loading...