Polish LLM Provider Usage

[joe223]

Thinking Mode Support by Provider

Whether reasoning/thinking mode can be disabled for major LLM providers and their models. All information below is verified against official provider documentation.

Last updated: 2026-04-10

Quick Reference

Provider	Model	Thinking Default	Can Disable?	API Parameter	Official Source
OpenAI	o1, o1-mini, o1-pro	Always on	❌ No	reasoning_effort: low/medium/high (reduce only)	platform.openai.com/docs/guides/reasoning
	o3, o3-mini	Always on	❌ No	reasoning_effort: low/medium/high
	GPT-5.4, GPT-5.3-codex, etc.	Varies	✅ Yes	reasoning_effort: none/low/medium/high/xhigh
	GPT-4o	No thinking	N/A	—
Anthropic	Claude Opus 4.6, Sonnet 4.6	Adaptive (on)	✅ Yes	thinking: {"type": "disabled"}	docs.anthropic.com/en/docs/build-with-claude/extended-thinking
	Claude Opus 4.5, Sonnet 4.5, Haiku	Off by default	✅ Yes	thinking: {"type": "disabled"}
	Claude Opus 4, Sonnet 4	Off by default	✅ Yes	thinking: {"type": "disabled"}
	Claude Mythos Preview	Adaptive (on)	❌ Not supported	—
Google	Gemini 2.5 Pro	Always on	❌ No	Cannot disable; min budget 128 tokens	ai.google.dev/gemini-api/docs/thinking
	Gemini 2.5 Flash	Dynamic (on)	✅ Yes	thinkingConfig.thinkingBudget: 0
	Gemini 2.5 Flash Lite	No thinking	N/A	—
	Gemini 3.1 Pro	Always on	❌ No	thinkingLevel: "minimal" not supported
	Gemini 3.1 Flash, Flash-Lite	Dynamic (on)	✅ Yes	thinkingLevel: "low"
DeepSeek	deepseek-reasoner (R1)	Always on	❌ No	Dedicated endpoint	api-docs.deepseek.com/guides/thinking_mode
	deepseek-chat (V3.2)	Off by default	✅ Yes	extra_body={"thinking": {"type": "enabled"}} to opt-in
Zhipu (智谱)	GLM-5, GLM-4.7	Always on	❌ No	thinking: {"type": "disabled"} accepted but forced by model design	docs.bigmodel.cn/cn/guide/capabilities/thinking
	GLM-4.5, GLM-4.6	Dynamic (on)	✅ Yes	thinking: {"type": "disabled"}
Qwen (通义千问)	QwQ-Plus, QwQ-32B	Always on	❌ No	Thinking-only model	www.alibabacloud.com/help/en/model-studio/qwq
	Qwen3 (hybrid variants)	Mixed*	✅ Yes	extra_body={"enable_thinking": false}
	Qwen3 (thinking-only variants)	Always on	❌ No	Model name ends with -thinking
	Qwen2.5, Qwen2	No thinking	N/A	—
ByteDance (豆包)	Doubao-1.5-pro, 1.5-lite	No thinking	N/A	—	www.volcengine.com/docs/6492/2165101
	Doubao-1.5-thinking-pro	Always on	✅ Yes	thinking: {"type": "disabled"}
	Doubao-Seed-1.6, 1.6-flash	Dynamic (auto)	✅ Yes	thinking: {"type": "disabled"}
	Doubao-Seed-1.6-thinking	Always on	✅ Yes	thinking: {"type": "disabled"}
MiniMax	M2.7, M2.7-highspeed	Always on	❌ No	thinking: {"type": "disabled"} is accepted but ignored	platform.minimax.io/docs/guides/text-m2-function-call
	M2.5, M2.5-highspeed	Always on	❌ No	Same — cannot disable

• Qwen3 hybrid defaults vary: Qwen3.5 series on by default; Qwen3, Qwen3-VL, Qwen3-Omni-Flash off by default. All hybrid models can toggle.

API Parameter Details

OpenAI

Official docs: https://platform.openai.com/docs/guides/reasoning

# o-series: can only reduce effort, cannot fully disable
response = client.chat.completions.create(
    model="o3",
    messages=[...],
    reasoning_effort="low"  # low / medium / high only
)

# GPT-5 series: can also set to "none" to disable
response = client.chat.completions.create(
    model="gpt-5.4",
    messages=[...],
    reasoning_effort="none"  # none / minimal / low / medium / high / xhigh
)

Anthropic

Official docs: https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking

# Disable thinking (all supported models except Mythos Preview)
response = client.messages.create(
    model="claude-sonnet-4-6",
    thinking={"type": "disabled"},
    max_tokens=1024,
    messages=[...]
)

# Adaptive thinking (Opus 4.6, Sonnet 4.6 — recommended)
thinking={"type": "adaptive", "effort": "high"}  # low / medium / high

# Manual mode (Opus 4.5, Sonnet 4.5, Haiku — deprecated on 4.6)
thinking={"type": "enabled", "budget_tokens": 10000}

Note: Claude Mythos Preview does NOT support thinking: {"type": "disabled"}.

Google Gemini

Official docs: https://ai.google.dev/gemini-api/docs/thinking

# Gemini 2.5 series — use thinkingBudget
from google.genai import types
config=types.GenerateContentConfig(
    thinking_config=types.ThinkingConfig(thinking_budget=0)  # disable
    # thinking_budget=-1  dynamic (default)
    # thinking_budget=1024  fixed budget
)

Model	Can Disable?	Parameter	Range
Gemini 2.5 Pro	❌ No	thinkingBudget	128–32768 (cannot set 0)
Gemini 2.5 Flash	✅ Yes	thinkingBudget: 0	0–24576
Gemini 2.5 Flash Lite	N/A	No thinking	—
Gemini 3.1 Pro	❌ No	thinkingLevel: "minimal" not supported	—
Gemini 3.1 Flash/Flash-Lite	✅ Yes	thinkingLevel: "low"	minimal/low/medium/high

DeepSeek

Official docs: https://api-docs.deepseek.com/guides/thinking_mode

# R1 — always thinks, dedicated reasoning endpoint
response = client.chat.completions.create(
    model="deepseek-reasoner",
    messages=[...]
)

# V3.2 — thinking off by default; enable with extra_body
response = client.chat.completions.create(
    model="deepseek-chat",
    messages=[...],
    extra_body={"thinking": {"type": "enabled"}}
)

Zhipu (智谱)

Official docs: https://docs.bigmodel.cn/cn/guide/capabilities/thinking

# GLM-4.5, GLM-4.6 — can disable
response = client.chat.completions.create(
    model="glm-4.6",
    messages=[...],
    extra_body={"thinking": {"type": "disabled"}}
)

# GLM-5, GLM-4.7 — cannot disable
# Official docs: "glm-5 glm-4.7 glm-4.5v 为强制思考" (forced thinking)
# sending {"type": "disabled"} is accepted but model still thinks

GLM-5 and GLM-4.7: Official documentation explicitly states these are "forced deep thinking" (强制思考). The thinking.type: disabled parameter is listed in the API schema but the models always produce thinking content. This mirrors MiniMax's behavior.

Qwen (通义千问)

Official docs: https://www.alibabacloud.com/help/en/model-studio/qwq

# Hybrid models — can disable
response = client.chat.completions.create(
    model="qwen3-32b",
    messages=[...],
    extra_body={"enable_thinking": False}
)

# With budget (DashScope / vLLM / SGLang deployments)
extra_body={"enable_thinking": True, "thinking_budget": 500}

Model	Thinking Default	Can Disable?
QwQ-Plus, QwQ-32B	Always on	❌ No (thinking-only)
Qwen3.5-plus, qwen3.5-flash	On by default	✅ Yes
Qwen3, Qwen3-VL, Qwen3-Omni-Flash	Off by default	✅ Yes
Qwen3-Omni-Turbo	No thinking	N/A
Model names ending in -thinking	Always on	❌ No (thinking-only)

ByteDance (豆包)

Official docs: https://www.volcengine.com/docs/6492/2165101

# OpenAI-compatible API — disable thinking
response = client.chat.completions.create(
    model="doubao-seed-1-6-251015",
    messages=[...],
    extra_body={"thinking": {"type": "disabled"}}
)
# Values: "enabled" / "disabled" / "auto"

# LAS/Daft API — thinking_type parameter
ArkLLMThinkingVision,
construct_args={
    "model": "doubao-seed-1.6",
    "thinking_type": "disabled",  # enabled / disabled / auto
}

MiniMax

Official docs: https://platform.minimax.io/docs/guides/text-m2-function-call

# WARNING: parameter is accepted but IGNORED
# MiniMax models always return thinking content regardless
response = client.messages.create(
    model="MiniMax-M2.7-highspeed",
    thinking={"type": "disabled"},  # accepted but ineffective
    max_tokens=1024,
    messages=[...]
)