5.21 Thinking Depth Configuration

Treat "thinking depth" like gears: shallow for speed, deep for stability.

What You'll Learn

• Set thinking budget for individual models
• Understand how the "variants" mechanism controls thinking depth
• Use Ctrl+T to switch between different thinking depths

---

Your Current Dilemma

• Same model, sometimes you want fast, sometimes deep, but don't know how to switch
• After writing config in opencode.json, unsure if it took effect
• When using relay/proxy models, unsure if you can still control thinking depth

---

When to Use This

• When you need: make "thinking depth" a switchable gear
• And don't want: change models or modify config every time

---

🎒 Before You Start

• [ ] Completed 5.1 Configuration Complete Guide
• [ ] Can start OpenCode normally

---

Core Concept

1. OpenCode uses model variants to save different thinking depths
2. Variants are model-level configurations with higher priority than defaults
3. Ctrl+T cycles through variants

ℹ️ What is "Thinking Depth"?

It refers to the model's "available thinking budget", such as Anthropic's thinking.budgetTokens. Higher values mean more tokens available for reasoning, but slower response and higher cost.

---

Follow Along

Step 1: Confirm Model Supports Thinking Variants

Why Not all models have variants, OpenCode checks capabilities.reasoning first.

How Choose a model that supports reasoning (e.g., Anthropic / Gemini 3 / OpenAI).

You Should See Variants like high / max appear in the model list.

---

Step 2: Set Thinking Budget for Individual Models in opencode.json

Why Variant config goes under provider.models.[modelID].variants, which can override defaults.

How Fill in the corresponding fields based on your Provider:

Anthropic Example (thinking.budgetTokens)

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "anthropic": {
      "models": {
        "claude-sonnet-4-5": {
          "variants": {
            "high": {
              "thinking": { "type": "enabled", "budgetTokens": 20000 }
            },
            "max": {
              "thinking": { "type": "enabled", "budgetTokens": 32000 }
            }
          }
        }
      }
    }
  }
}

Gemini 3 Example (thinkingConfig.thinkingBudget)

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "google": {
      "models": {
        "gemini-3-flash": {
          "variants": {
            "high": {
              "thinkingConfig": { "includeThoughts": true, "thinkingBudget": 16000 }
            },
            "max": {
              "thinkingConfig": { "includeThoughts": true, "thinkingBudget": 24576 }
            }
          }
        }
      }
    }
  }
}

You Should See Variant values take effect after restart.

---

Step 3: Switch Thinking Depths with Ctrl+T

Why After setting up variants, using a shortcut to switch is more convenient.

How Press Ctrl+T in the chat input box to cycle:

(none) → high → max → (none) → high → ...

You Should See Status bar shows current variant name (e.g., high).

---

Step 4: Customize Variant Names (Optional)

Why Variant names aren't fixed, you can change them to things like "Deep Think" / "Fast".

How Use custom keys in variants:

{
  "provider": {
    "anthropic": {
      "models": {
        "claude-sonnet-4-5": {
          "variants": {
            "fast": { "thinking": { "type": "enabled", "budgetTokens": 8000 } },
            "deep": { "thinking": { "type": "enabled", "budgetTokens": 32000 } }
          }
        }
      }
    }
  }
}

You Should See Ctrl+T switches between "fast/deep".

ℹ️ Custom Variants Are "Added", Not "Replaced"

OpenCode merges the variants you write in opencode.json with default variants. If you only want custom variants, explicitly disable the default high/max:

{
  "provider": {
    "anthropic": {
      "models": {
        "claude-sonnet-4-5": {
          "variants": {
            "high": { "disabled": true },
            "max": { "disabled": true },
            "fast": { "thinking": { "type": "enabled", "budgetTokens": 8000 } },
            "deep": { "thinking": { "type": "enabled", "budgetTokens": 32000 } }
          }
        }
      }
    }
  }
}

How to Configure Third-Party Relays

If your relay uses openai-compatible, it defaults to reasoningEffort. Example:

{
  "provider": {
    "relay": {
      "options": {
        "baseURL": "https://your-relay.example.com/v1",
        "apiKey": "{env:RELAY_API_KEY}"
      },
      "models": {
        "gpt-5": {
          "variants": {
            "low": { "reasoningEffort": "low" },
            "high": { "reasoningEffort": "high" }
          }
        }
      }
    }
  }
}

If your relay actually forwards to Anthropic (still using openai-compatible SDK), you can directly override with Anthropic fields:

{
  "provider": {
    "relay": {
      "options": {
        "baseURL": "https://your-relay.example.com/v1",
        "apiKey": "{env:RELAY_API_KEY}"
      },
      "models": {
        "claude-sonnet-4-5": {
          "variants": {
            "high": {
              "thinking": { "type": "enabled", "budgetTokens": 20000 }
            },
            "max": {
              "thinking": { "type": "enabled", "budgetTokens": 32000 }
            }
          }
        }
      }
    }
  }
}

Prerequisite: your relay server forwards the thinking field to Anthropic as-is.

---

Checklist ✅

• [ ] opencode.json contains provider.models.[modelID].variants
• [ ] Variant name appears in status bar after startup
• [ ] Ctrl+T can cycle through variants

---

Common Issues

Symptom	Cause	Solution
Ctrl+T does nothing	Current model has no variants	Switch to a reasoning-capable model or add variants
Variants exist but don't show	Haven't switched to a variant yet	Press Ctrl+T once
Config doesn't take effect	Wrong model ID	Copy complete ID from model list
Relay shows no change	Using openai-compatible, only supports reasoningEffort	Manually override parameters in variants

---

Lesson Summary

You learned:

1. Variants are "thinking depth gears", configured in provider.models.[modelID].variants
2. Default variants are auto-generated by ProviderTransform, can be overridden by config
3. Ctrl+T cycles through variants

---

Next Lesson Preview

Next we'll learn Debugging & Diagnostic Tools.

You'll learn:

• How to use opencode debug commands
• Diagnose LSP, config, and search issues
• Analyze OpenCode like a developer

---

Appendix: Source Code Reference

Click to expand source code locations

Updated: 2026-01-16

Feature	File Path	Lines
Variant generation entry	src/provider/transform.ts	297-477
Reasoning filter & exclusion	src/provider/transform.ts	298-301
Anthropic thinking budget defaults	src/provider/transform.ts	371-385
Gemini 3 thinking budget defaults	src/provider/transform.ts	421-439
Variant config Schema	src/config/config.ts	818-833
Variant config merge	src/provider/provider.ts	929-936
Ctrl+T default shortcut	src/config/config.ts	632-688
Ctrl+T command binding	src/cli/cmd/tui/app.tsx	393-399
Variant cycling logic	src/cli/cmd/tui/context/local.tsx	310-346
Variant display logic	src/cli/cmd/tui/component/prompt/index.tsx	696-700
Variant name rendering	src/cli/cmd/tui/component/prompt/index.tsx	946-950
Variant applied to LLM params	src/session/llm.ts	96-109
Variant keybind config	src/config/config.ts	632-688

Key Constants:

• WIDELY_SUPPORTED_EFFORTS = ["low", "medium", "high"]
• OPENAI_EFFORTS = ["none", "minimal", "low", "medium", "high", "xhigh"]