OpenAI’s Ted Sanders Explains Model Sizing and Version Numbering

OpenAI’s Ted Sanders recently clarified how the company thinks about model sizing and version numbering conventions, offering insight into the factors that determine both model performance and naming schemes.

Understanding Model Size Trade-offs

According to Sanders, LLM performance operates across multiple dimensions without guaranteed consistency. A model excelling at mathematics won’t necessarily perform best at generating jokes. However, despite this multidimensional complexity, a single variable explains much of the performance variation: model size.

Sanders outlined three core characteristics that scale with model size. Larger models demonstrate greater performance and reliability across tasks. They cost more per token to operate. And they exhibit better recall of esoteric facts - the kind of obscure information that smaller models might miss.

This sizing principle applies to OpenAI’s current model lineup, which Sanders confirmed follows a clear hierarchy: Luna < Terra < Sol < Astra. Each step up in this ordering represents a larger model with the corresponding trade-offs between capability and cost.

Decoding Version Numbers

Sanders also explained the logic behind version numbering, which has become more complex as model releases have accelerated. Modern version numbers communicate three pieces of information simultaneously: release order, relationships to other releases, and the magnitude of improvement over previous versions.

Using a hypothetical example, Sanders noted that a 6.1 release would indicate several things. First, it launched after 5.6 in chronological order. Second, it shares similarities with other 6.x models in terms of training data or personality - which is why it wouldn’t be numbered 5.7 instead. Third, the incremental version bump signals that the improvements, while meaningful, don’t represent the kind of breakthrough that would justify jumping to version 7 with its accompanying headlines and elevated expectations.

Balancing Capability Against Cost

The sizing framework Sanders described creates a practical decision tree for developers. Applications requiring maximum accuracy or handling of obscure knowledge domains benefit from larger models like Astra, accepting the higher per-token costs. Conversely, high-volume applications with simpler requirements can achieve better economics by selecting smaller models like Luna or Terra.

This trade-off becomes particularly important when designing systems that make multiple LLM calls. A common pattern involves using a smaller model for initial filtering or classification, then routing only selected requests to a larger, more expensive model for final processing.

Version Signaling and User Expectations

The version numbering explanation reveals how OpenAI manages user expectations through naming conventions. Major version bumps (like moving from 6.x to 7.0) signal transformative improvements that warrant industry attention. Minor version increments indicate meaningful progress within an established model family, suggesting refinements rather than fundamental architectural changes.

This approach helps developers understand which upgrades require thorough retesting of their applications versus which represent incremental improvements that maintain backward compatibility in behavior and personality.

Sanders’ comments, shared on October 28th, 2026, provide rare transparency into the reasoning behind model development and release strategies. Understanding these principles helps developers make informed choices about which models to deploy and when to invest effort in migrating to newer versions.

The multidimensional nature of LLM performance means that bigger isn’t always better for every use case - but the consistent relationship between size, capability, and cost provides a reliable framework for making architectural decisions.