OpenCode reasoning effort

Reasoning and thinking models bill extra output tokens for internal chains before the visible reply. OpenCode exposes provider options like reasoningEffort and budgetTokens per model and per agent.

This page answers opencode change reasoning effort, opencode no thinking, and custom provider thinking options.

Hub: Tokenminning in OpenCode.

Change reasoning effort

Configure at provider model level and override per agent in opencode.json (Models ):


{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "anthropic": {
      "models": {
        "claude-sonnet-4-5": {
          "options": {
            "reasoningEffort": "low"
          }
        }
      }
    }
  },
  "agent": {
    "plan": {
      "model": "anthropic/claude-haiku-4-20250514"
    },
    "build": {
      "model": "anthropic/claude-sonnet-4-5",
      "options": {
        "reasoningEffort": "medium"
      }
    }
  }
}

Setting	Token effect
`reasoningEffort`	Higher → more thinking tokens before answer
`budgetTokens`	Caps reasoning budget when provider supports it
Thinking model on Plan	Overhead on questions Haiku could handle

Turn off thinking

To disable thinking on a route:

Pick a non-reasoning model ID (no -thinking / extended-reasoning variant)
Set lowest reasoningEffort your provider allows
Do not set thinking models as the global model default

For opencode no thinking on Plan specifically, assign Haiku or a non-reasoning mid-tier to agent.plan only.

Thinking budgets (custom providers)

OpenRouter and custom OpenAI-compatible gateways often expose thinking controls via provider options. Example pattern:


{
  "provider": {
    "openrouter": {
      "models": {
        "anthropic/claude-sonnet-4-5": {
          "options": {
            "reasoningEffort": "low",
            "budgetTokens": 8000
          }
        }
      }
    }
  }
}

Exact keys depend on the provider — verify in Providers and your gateway docs. Wrong options silently fall back to provider defaults.

When reasoning is worth it

Task	Model
Grep-style questions, plan review	Haiku / mid-tier, no thinking
Multi-file refactor	Capable model, moderate effort
Novel architecture, deep debug	Thinking variant, scoped prompt

Use Plan with cheap models; escalate effort only on Build after you approve the approach.

Subagents and parallel thinking

@general, @scout, and Task-delegated subagents spawn sessions with their own context and model choices. Restrict with permission.task; prefer read-only @explore for search.

Each parallel subagent can run thinking models independently — multiply overhead fast.

Troubleshooting

High output on simple tasks — Thinking default on Plan. Switch Plan to Haiku; lower reasoningEffort.

Effort change ignored — Options must match provider + model ID in config. Check Providers merge order.

Stats spike after model change — New conversation when switching models; cache behavior resets.