Thinking and reasoning effort

Reasoning controls

Choose an effort level

low

Faster responses and lighter reasoning.

medium

Default. Balanced speed and reasoning depth.

high

More reasoning for difficult analysis and coding.

xhigh

Deep reasoning for complex architecture and long-horizon agent work.

max

Highest verified GPT-5.6 reasoning level. Expect the most latency and token use.

Do not send reasoning_effort: "fast". fast is a service-tier alias, not a thinking level. For priority serving, use service_tier: "priority" separately.

Request examples

Set effort in the request body

Use the top-level reasoning_effort field with Chat Completions. With the Responses API, use the canonical reasoning.effort object. Replace the base URL with https://2.jatevo.aifor $JTVO-holder inference.

Chat Completions · API Master

curl https://api.jatevo.ai/v1/chat/completions \
  -H "Authorization: Bearer $JATEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "messages": [{"role": "user", "content": "Solve this carefully."}],
    "reasoning_effort": "high"
  }'

Responses API · API Master

curl https://api.jatevo.ai/v1/responses \
  -H "Authorization: Bearer $JATEVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "input": "Solve this carefully.",
    "reasoning": {"effort": "high"}
  }'

Responses API · $JTVO holder endpoint

curl https://2.jatevo.ai/v1/responses \
  -H "Authorization: Bearer $JTVO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "input": "Solve this carefully.",
    "reasoning": {"effort": "xhigh"}
  }'

Agent clients

Configure thinking once per model

OpenClaw provider setup

{
  "agents": {
    "defaults": {
      "model": { "primary": "jatevo/gpt-5.6-sol" },
      "models": { "jatevo/gpt-5.6-sol": {} }
    }
  },
  "models": {
    "providers": {
      "jatevo": {
        "baseUrl": "https://api.jatevo.ai/v1",
        "apiKey": "YOUR_JATEVO_API_KEY",
        "api": "openai-responses",
        "models": [
          {
            "id": "gpt-5.6-sol",
            "name": "GPT-5.6 Sol",
            "reasoning": true,
            "input": ["text", "image"],
            "contextWindow": 372000,
            "maxTokens": 131072,
            "compat": {
              "supportedReasoningEfforts": ["low", "medium", "high", "xhigh", "max"],
              "reasoningEffortMap": { "xhigh": "xhigh", "max": "max" }
            }
          }
        ]
      }
    }
  }
}

After adding the model, select the effort for the current OpenClaw session with /think low, /think medium, /think high, or /think xhigh. The variants object shown in some client examples is not an OpenClaw model-provider field.

OpenCode

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "jatevo": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://api.jatevo.ai/v1",
        "apiKey": "{env:JATEVO_API_KEY}"
      },
      "models": {
        "gpt-5.6-sol": {
          "name": "GPT-5.6 Sol",
          "reasoning": true,
          "options": {
            "reasoningEffort": "high",
            "reasoningSummary": "detailed"
          }
        }
      }
    }
  }
}

Codex CLI · $JTVO endpoint

model = "gpt-5.6-sol"
model_reasoning_effort = "xhigh"
model_provider = "jatevo"

[model_providers.jatevo]
name = "openai"
base_url = "https://2.jatevo.ai/backend-api/codex"
wire_api = "responses"
requires_openai_auth = true

Client configuration names are camelCase in OpenCode (reasoningEffort), while the raw Chat Completions API field is snake_case (reasoning_effort). OpenClaw exposes configured levels through /think.

Important behavior

Defaults, compatibility, and cost

Defaults are model-specific

GPT-5.6 Sol defaults to low; Terra and Luna default to medium.

The user can override it

Set effort per request or configure a persistent model variant in the client.

Support is model-specific

These controls target reasoning-capable models. Jatevo may normalize unsupported effort levels for a specific upstream model.

Rate is not total cost

The published per-token rate does not change by effort, but higher effort can generate more reasoning/output tokens and increase total usage.

Provider note: Kimi K2.7 defaults to medium on the Jatevo route and supports low, medium, and high. Kimi K3 is currently forced to its supported maximum mode. The GPT examples above should not be copied blindly to every open model.

Need a model-specific example?

Browse providers