[ ABORT TO HUD ]
SEQ. 1
Reasoning Models Architecture & Controls
OpenAI Reasoning Models: o1 & o3-mini
OpenAI's reasoning models (o1 and o3-mini) are trained with reinforcement learning to generate internal chains of thought before producing a response. This allows them to excel in STEM, complex coding, and multi-step analytical reasoning.
Calibrating Reasoning Effort
On o3-mini, you can calibrate how many reasoning tokens the model uses via the reasoning_effort parameter:
| Effort Level | Latency | Best For |
|---|---|---|
low | Fastest | Straightforward code fixes, basic math, fast decision routing |
medium (default) | Balanced | General software engineering, data analysis, standard reasoning |
high | Thorough | Complex mathematical proofs, competitive programming, intricate architecture design |
const response = await openai.chat.completions.create({
model: "o3-mini",
reasoning_effort: "high",
messages: [
{ role: "developer", content: "You are a senior algorithmic engineer." },
{ role: "user", content: "Optimize this dynamic programming solution for memory bounds." }
]
});Developer Messages
Reasoning models replace the legacy system role with the developer role, ensuring clear instruction hierarchy while preserving the model's safety and reasoning boundaries.
💡 Best Practice: Reasoning models natively support Structured Outputs (
strict: true) and Function Calling, making them powerful reasoning engines for agentic workflows.SYNAPSE VERIFICATION
QUERY 1 // 3
Which parameter controls the depth of reasoning tokens on o3-mini?
temperature
reasoning_effort ('low' | 'medium' | 'high')
max_reasoning_tokens
thinking_depth