Reasoning Effort

Configuring thinking token budgets and reasoning effort across GPT-6, DeepSeek, and Claude models.

Modern reasoning models generate internal chains of thought prior to emitting a final response. Toknix allows developers to tune reasoning intensity to balance depth against response latency.

Reasoning Mechanics

  • Reasoning Tokens: Tokens produced during the internal reasoning phase count toward the model's output token limit and are billed at the model's standard output rate.
  • Supported Models: GPT-6 Sol, GPT-5.6 Sol, DeepSeek V4 Pro, Kimi K3, and Claude Opus 5.5.

OpenAI Configuration

Pass the reasoning_effort parameter:

json
{
  "model": "gpt-6-sol",
  "reasoning_effort": "high",
  "messages": [
    {"role": "user", "content": "Formulate a formal proof of Byzantine Fault Tolerance."}
  ]
}

Supported values: "low", "medium", "high".

Anthropic Configuration

On Anthropic format requests, supply the thinking object:

json
{
  "model": "claude-opus-5-5-max",
  "max_tokens": 16000,
  "thinking": {
    "type": "enabled",
    "budget_tokens": 4096
  },
  "messages": [
    {"role": "user", "content": "Analyze potential race conditions in this concurrency loop."}
  ]
}