Skip to content
FirstPrinciples Help Center

Model, reasoning mode, and Council mode

The chat composer sits at the bottom of Central Chat. It is where you enter a request and, when needed, choose the model, reasoning mode, or Council.

Supported models

  • OpenAI — GPT 5.4 Nano, GPT 5.4 Thinking, and GPT 5.5 Max.

  • Anthropic — Claude Opus 5 and Claude Sonnet 5.

  • GLM 5.2 — An open-weight model post-trained by FirstPrinciples for physics and mathematics. This is a text only model, not multimodal.

  • Kimi K2— An open-weight model post-trained by FirstPrinciples for physics and mathematics. This is a text only model, not multimodal.

How model selection works

  • By default, Theo uses OpenAI Fast with GPT 5.4 Nano. You can choose any other available model when a different capability, perspective, or reasoning depth is more useful.

  • Depending on the question, the selected model can call the appropriate agents, subagents, and tools needed to complete the work.

  • Speed or reasoning mode lets you use a faster option for exploration or a more intensive option for difficult reasoning and verification.

Theo is designed to aggregate frontier models and fine-tuned open-weight models in one research environment. This gives researchers control over which model is most useful for a task while keeping access to Theo’s phases, sources, artifacts, and research context.

The available model lineup will be updated as the industry advances and as FirstPrinciples improves its own research models.

Council

  • Use Council for specific, difficult questions that benefit from multiple perspectives, independent approaches, or debate.

  • Council requires at least two models.

  • Council models receive the relevant chat history and artifacts produced by the system, allowing them to reason from the shared research context.

  • Choose an OpenAI model and an Anthropic model when you want two frontier-model families to debate an issue or compare conclusions.

  • Add the FirstPrinciples-tuned GLM and Kimi models when you want to compare their reasoning traces and see how the models are approaching the problem.

Council and more intensive reasoning modes may take longer or use more workspace resources than the default fast model.