Chain of Thought & Reasoning Architecture — Reasoning Models

A New Model in AI

The reasoning model lineage (from early o1/o3-mini previews to today's unified GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna family) represents a fundamental shift. Instead of generating answers token-by-token immediately, these models use reinforcement learning to generate a hidden Chain of Thought (CoT)—billed at output token rates—before producing the final output.

The Evolution & Production Status (October 2026)

How Reasoning Models Think

1. Break the problem into smaller steps.
2. Try different approaches.
3. Recognize mistakes and backtrack.
4. Synthesize a final, accurate answer.

Prompting Reasoning Models

• Keep it simple: State the problem directly. Do NOT say "think step by step."
• Provide edge cases: Give constraints the model should consider.
• Use the developer role: Reasoning models use developer instead of system.

// Reasoning models use the "developer" role and reasoning.effort:
const response = await openai.responses.create({
  model: "gpt-6.1-sol",
  reasoning: { effort: "high" },  // low | medium | high | xhigh | max
  input: [
    { role: "developer", content: "You are a math olympiad judge. Be rigorous." },
    { role: "user", content: "Prove that sqrt(2) is irrational." }
  ]
});