OpenAI’s Ted Sanders shares how to pick the right GPT model

Performance beyond model size

According to The Decoder, OpenAI’s Ted Sanders explained that the performance of its language models is multidimensional. A model that excels in one area, such as mathematical reasoning, may not perform as well in another, like generating creative content. While model size often influences performance gaps, newer and smaller models can outperform older, larger ones due to improvements in training data or architecture.

How OpenAI categorizes its models

Sanders described the challenges in simplifying OpenAI’s model naming conventions, as dimensions like model size, reasoning effort, and multi-agent capabilities are distinct. Reducing these dimensions to a single metric, such as an "IQ score," would obscure critical differences in performance and use cases.

He highlighted GPT-6.1 Terra as an example. While it outperforms the larger Astra model in technical communication and matches it in image processing and computer use at a lower cost, it falls short in most other areas.

Rule of thumb for model selection

Sanders offered a straightforward guideline for choosing among OpenAI’s GPT models: "If you care about quality only, pick Astra. If you're a bit price sensitive, pick Sol. If you're very price sensitive, pick Terra or Luna."

What this means

Lazyfounder analysis — our interpretation, not reported fact.

Sanders’ insights clarify how OpenAI positions its models for different operational and budgetary needs. For founders and operators, this framework removes some of the guesswork in selecting a model, particularly when balancing cost and performance. The emphasis on multidimensional performance also underscores that the most expensive model may not always be the best fit—an important consideration for startups optimizing for specific tasks like technical communication or multi-agent systems.

Related on Lazyfounder

• OpenAI
• More ai news
• OpenAI’s AI model claims disproof of Exponential Time Hypothesis