Reasoning Effort & Adaptive Thinking — Reasoning Models
[ ABORT TO HUD ]
SEQ. 1
SEQ. 2
Reasoning Effort & Adaptive Thinking
Calibrating Reasoning Depth
The reasoning.effort parameter lets you control how much compute the model spends thinking internally before generating visible tokens. Remember that hidden reasoning tokens are billed at the model's standard output rate.
GPT-6 Effort Levels
| Level | Model Support | Cost & Latency | Best For |
|---|---|---|---|
| none | gpt-6-luna only | Ultra-low (~0.9s TTFT) | High-volume extraction, formatting, fast chat routing |
| low | All GPT-6 models | Low | Light code edits, syntax translation, quick classification |
| medium (default) | All GPT-6 models | Moderate | General multi-step logic, standard software engineering |
| high / xhigh | Sol & Astra (and Luna) | High | Multi-file architectural refactoring, rigorous math, complex debugging |
| max | gpt-6.1-sol & gpt-6-astra | Maximum | Long-horizon agentic workflows, formal verification, scientific research |
Model-Specific Effort Behavior
While GPT-6 Luna explicitly supports effort: "none" for non-reasoning speed (making it the ideal drop-in successor to deprecated gpt-4o-mini), GPT-6.1 Sol requires at least effort: "low" and rejects unreasoned configurations. Changing reasoning.effort across turns preserves prompt cache prefixes.
// Non-reasoning fast classification on GPT-6 Luna:
const fast = await openai.responses.create({
model: "gpt-6-luna",
reasoning: { effort: "none" },
input: "Classify sentiment: Great product launch!"
});
// Deep architectural reasoning on GPT-6.1 Sol:
const complex = await openai.responses.create({
model: "gpt-6.1-sol",
reasoning: { effort: "high" },
input: "Design a distributed consensus algorithm for a 10-node cluster"
});
💡 Cost Tip: Keep GPT-6 models on the default
medium effort for everyday coding, drop to none or low on gpt-6-luna for high-throughput routing, and reserve high / xhigh / max on gpt-6.1-sol or gpt-6-astra for genuinely hard problems.⌨ HANDS-ON LABCalibrate Reasoning Effort Levels
⭐ +200 XPControl reasoning budget and inference latency on GPT-6.1 Sol. Submit a Responses API request with reasoning.effort set to high.
1POST to /v1/responses with reasoning.effort set to high on gpt-6.1-sol.
OBJECTIVE 1 / 1 — type "hint" if stuck
SYNAPSE VERIFICATION
QUERY 1 // 3
What does the reasoning.effort parameter control?
Model temperature
How much time and compute the model spends thinking internally before answering
The maximum input window size
Network encryption