Reasoning - MyTokenGate

1. Overview

Reasoning models are AI systems based on deep learning that solve complex tasks through logical deduction, knowledge association, and context analysis. Typical applications include mathematical problem solving, code generation, logical judgment, and multi-step reasoning scenarios.

2. Current Support

MyTokenGate currently offers models with strong reasoning capabilities:

Claude Series

claude-fable-5 - Frontier, always-on adaptive thinking
claude-sonnet-5 - Balanced default, high throughput
claude-opus-4-8 - Previous frontier, long-horizon reasoning

GPT Series

gpt-5.6-sol - Flagship reasoning
gpt-5.4 - Everyday workhorse reasoning
gpt-5.5 - Previous-generation flagship

Gemini Series

gemini-3.1-pro-preview - Complex multimodal reasoning

Open-weights reasoning (OpenAI-compatible reasoning_content)

glm-5.1 / glm-5.2 - GLM reasoning models
deepseek-v4-pro - Deep reasoning flagship (coming soon)

3. Usage Example


from openai import OpenAI
 
client = OpenAI(
    base_url='https://gateway.mytokengate.com/v1',
    api_key='your-api-key'
)
 
response = client.chat.completions.create(
    model="claude-sonnet-5",
    messages=[
        {"role": "user", "content": "Analyze step by step: If all A are B, and all B are C, then are all A C?"}
    ],
    max_tokens=4096,
    stream=True
)
 
for chunk in response:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

4. Reasoning Controls (Effort and Thinking Budget)

Each vendor controls “thinking” depth with different parameters and value ranges:

GPT-5 series — reasoning.effort, values minimal/low/medium (default)/high, with none/xhigh/max on newer models; minimal means near-zero thinking and lowest latency. Reasoning tokens are billed as output.
Claude extended thinking — thinking.type="enabled" + budget_tokens (thinking token budget); on Claude 4.6 this explicit mode is deprecated but still works, and frontier models use adaptive thinking by default.
Gemini — thinkingConfig with thinkingBudget (0 off / -1 dynamic) and the newer thinkingLevel; Gemini 3 Pro cannot fully disable thinking.
Open-weights reasoning (GLM / DeepSeek) — over the OpenAI-compatible interface, thinking is returned in the reasoning_content field and must be preserved and replayed verbatim in tool-calling flows; see Interleaved Thinking.

MyTokenGate passes these parameters through its unified gateway; refer to each model page and the upstream docs for exact values.

5. Best Practices

Break down complex tasks: Split complex problems into multiple steps
Clear instructions: Provide clear context and requirements
Stream output: Use streaming for long reasoning tasks

6. Notes

Ensure you use the correct API key for authentication
Different models have different context window limits
Check Models for the latest supported models