Why GPT-5.6 was followed by 6.1 rather than 5.7: an OpenAI employee decodes the model names

In a Hacker News thread about the release of GPT-6.1 Terra, OpenAI's Ted Sanders laid out how the company names its models. Simon Willison quoted him on his blog on October 28. The comment gives developers a simple key to a lineup that now has four sizes and several generations side by side.

The name tells you the size, the version number tells you three other things

Sanders says language model performance is multidimensional, and the model that is best at math is not necessarily the best at jokes. Still, in his view much of the variation is explained by a single variable: model size. Bigger models tend to be more capable and reliable, cost more per token and recall obscure facts better. OpenAI's models are therefore ordered from smallest to largest: Luna, Terra, Sol, Astra. The Latin names run from the Moon through the Earth and the Sun to the stars.

According to Sanders, the version number says three things at once: the order of release, the relationship to other models and the size of the leap. Version 6.1 came out after 5.6 and shares some training data and personality with the other 6.1 models, which is why it is not called 5.7. It is also not a big enough jump to earn the headlines and expectations that come with a 7.

Sanders noted that OpenAI used to signal size with numbers as well. That is why GPT-4.1, an incremental update to GPT-4, shipped after GPT-4.5, which had been a big jump over GPT-4. He closed by asking whether anyone had a better naming idea.

Developers get a rule for picking models, exceptions included

Sanders offers practical advice. If you only care about quality, pick Astra. If you are somewhat price sensitive, pick Sol, and if you are very price sensitive, Terra or Luna. For a team budgeting an agent, that is a more useful guide than marketing pages. In the API, Terra costs $1 per million input tokens and $5 per million output tokens, half the price of Sol.

He immediately adds exceptions. Newer models may have better data or techniques, and then a smaller model beats a bigger one in some respects. He cites Sonnet 5.5 against Fable 5.1 and Gemini 3.8 Flash against 3.1 Pro. Terra itself, he says, is better than Sol and Astra at technical communication, similar on vision and computer use but much faster, and a bit weaker on most other things.

The most important part comes last. In Sanders's account, model size is a different axis from reasoning effort, and both differ from the number of agents deployed. If OpenAI folded them into one label like „more IQ“, users would lose important details. A developer is therefore choosing a combination of at least three settings, plus the model generation on top.

One forum comment does not replace documentation

Sanders is speaking as an OpenAI employee in a discussion thread, and this is not an official company document. His comparisons also rest on a vague „better in ways“. He does not say by how much or on which tasks, so a team cannot tell from them when it pays to step down a tier.

The size rule is only a starting point, as Sanders himself admits. If Terra beats Astra in one discipline, the Luna to Astra ladder is not enough, and teams will have to measure their own workloads anyway. The naming helps with orientation, but it will not make the decision for the developer.

The next release will test the rule

The simplest test is the next model. If OpenAI names its next step by the same rules and explains it in the documentation as clearly as Sanders did in a comment, the system holds. If another exception like GPT-4.1 after GPT-4.5 arrives, the key to the names will stay buried in forum threads.

The second signal is whether OpenAI publishes comparisons broken down by discipline. The claim that Terra beats Astra at technical communication needs numbers before anyone can buy on it.

Lilith's verdict

OpenAI arranged its models from the Moon to the stars and added a number that means three things at once. The developer budgeting an agent found the clearest map of that sky in a Hacker News comment, drawn on the company's behalf by a single employee.