Thinking (Reasoning) - Backboard API

Pass a thinking object on send_message / sendMessage to let the model reason before responding. Omit it or set null to disable. Pass {} for provider defaults.

thinking is a per-turn parameter — pass it on every call where you want reasoning enabled.

import asyncio
from backboard import BackboardClient

async def main():
    client = BackboardClient(api_key="YOUR_API_KEY")

    response = await client.send_message(
        "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
        llm_provider="openai",
        model_name="o3",
        thinking={"effort": "high"},
        stream=False
    )
    print(response.reasoning)
    print(response.content)

if __name__ == "__main__":
    asyncio.run(main())

Fields

Field	Type	Description
`effort`	string	Reasoning depth: `low`, `medium`, `high`, or `max`
`budget_tokens`	integer	Max tokens for internal reasoning
`max_tokens`	integer	Max tokens for reasoning output
`exclude_reasoning`	boolean	Omit reasoning traces from response

Provider Support

Provider	Supported Fields	Example
OpenAI	`effort`	`{"effort": "high"}`
Anthropic	`effort`, `budget_tokens`	`{"effort": "high"}`
AWS Bedrock	`effort`, `budget_tokens`	`{"effort": "high"}`
Google (Gemini 2.5)	`budget_tokens`	`{"budget_tokens": 8192}`
Google (Gemini 3)	`effort`	`{"effort": "high"}`
xAI	`effort`	`{"effort": "high"}`
OpenRouter	`max_tokens`, `exclude_reasoning`	`{"max_tokens": 4096}`
Cerebras	`max_tokens`, `exclude_reasoning`	`{"max_tokens": 4096}`

Provider Details

OpenAI: Models: o1, o3, o4, gpt-5 series. Default effort: medium.

Anthropic: Claude Opus 4.7 and newer (including Opus 5 and Fable 5) use adaptive thinking — pass effort (low, medium, high, max); budget_tokens is not accepted for these models. Older models (Claude 3.7 Sonnet, Sonnet 4, Opus 4–4.6) use budget_tokens (min 1,024). Backboard routes to the correct mode automatically.

AWS Bedrock: Bedrock-hosted Anthropic Claude models. Claude Opus 4.7 and newer (including Opus 5 and Fable 5) use adaptive thinking — pass effort; budget_tokens is rejected for these models. Older Bedrock Claude models accept budget_tokens (min 1,024) or effort. Backboard routes to the correct mode automatically.

Google (Gemini 2.5): Max budget: 32,768 (Pro) / 24,576 (others).

Google (Gemini 3): Uses effort (low through max). Thinking is on by default.

xAI: Reasoning models — grok-4.3 and its aliases grok-latest, grok-3*, grok-4*, grok-4-fast-reasoning, grok-4-1-fast-reasoning, grok-3-mini* — accept effort (low, medium, high, max); max maps to xAI’s highest level. Omit thinking to avoid setting a Backboard control (xAI’s own default is low); pass {} for Backboard’s default medium. The *-non-reasoning variants do not support effort.

OpenRouter / Cerebras: Use max_tokens (int > 0) and exclude_reasoning (bool).

Best for complex tasks: math, code generation, multi-step reasoning, analysis.

Thinking uses extra tokens and adds latency. Skip it for simple questions.

Non-Streaming Examples

OpenAI

Python
JavaScript
TypeScript

import asyncio
from backboard import BackboardClient

async def main():
    client = BackboardClient(api_key="YOUR_API_KEY")

    response = await client.send_message(
        "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
        llm_provider="openai",
        model_name="o3",
        thinking={"effort": "high"},
        stream=False
    )
    print(response.reasoning)
    print(response.content)

if __name__ == "__main__":
    asyncio.run(main())

Anthropic

Use effort for Claude Opus 4.7+ (adaptive thinking). Use budget_tokens for older models.

Python
JavaScript
TypeScript

# Claude Opus 4.7+ — use effort
response = await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="anthropic",
    model_name="claude-opus-4-7-20250501",
    thinking={"effort": "high"},
    stream=False
)
print(response.reasoning)
print(response.content)

AWS Bedrock

Use effort for Claude Opus 4.7+ on Bedrock. Use budget_tokens for older Bedrock Claude models.

Python
JavaScript
TypeScript

response = await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="aws-bedrock",
    model_name="anthropic.claude-opus-4-7",
    thinking={"effort": "high"},
    stream=False
)
print(response.reasoning)
print(response.content)

Google (Gemini 2.5)

Python
JavaScript
TypeScript

response = await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="google",
    model_name="gemini-2.5-pro",
    thinking={"budget_tokens": 8192},
    stream=False
)
print(response.reasoning)
print(response.content)

Google (Gemini 3)

Python
JavaScript
TypeScript

response = await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="google",
    model_name="gemini-3.1-pro-preview",
    thinking={"effort": "high"},
    stream=False
)
print(response.reasoning)
print(response.content)

xAI

Python
JavaScript
TypeScript

response = await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="xai",
    model_name="grok-4.3",
    thinking={"effort": "high"},
    stream=False
)
print(response.reasoning)
print(response.content)

OpenRouter

Python
JavaScript
TypeScript

response = await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="openrouter",
    model_name="minimax/minimax-m1",
    thinking={"max_tokens": 4096},
    stream=False
)
print(response.reasoning)
print(response.content)

Cerebras

Python
JavaScript
TypeScript

response = await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="cerebras",
    model_name="z-ai/glm-4.7",
    thinking={},
    stream=False
)
print(response.reasoning)
print(response.content)

Streaming Examples

Same as above but with stream=True. The model thinks first, then streams the response.

OpenAI

Python
JavaScript
TypeScript

async for chunk in await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="openai",
    model_name="o3",
    thinking={"effort": "high"},
    stream=True
):
    t = chunk.get("type")
    if t == "reasoning_streaming":
        print(chunk.get("content", ""), end="", flush=True)
    elif t == "reasoning_ended":
        print("\n---------Reasoning Ended, Response Started---------")
    elif t == "content_streaming":
        print(chunk.get("content", ""), end="", flush=True)

Anthropic

Python
JavaScript
TypeScript

async for chunk in await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="anthropic",
    model_name="claude-opus-4-7-20250501",
    thinking={"effort": "high"},
    stream=True
):
    t = chunk.get("type")
    if t == "reasoning_streaming":
        print(chunk.get("content", ""), end="", flush=True)
    elif t == "reasoning_ended":
        print("\n---------Reasoning Ended, Response Started---------")
    elif t == "content_streaming":
        print(chunk.get("content", ""), end="", flush=True)

AWS Bedrock

Python
JavaScript
TypeScript

async for chunk in await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="aws-bedrock",
    model_name="anthropic.claude-opus-4-7",
    thinking={"effort": "high"},
    stream=True
):
    t = chunk.get("type")
    if t == "reasoning_streaming":
        print(chunk.get("content", ""), end="", flush=True)
    elif t == "reasoning_ended":
        print("\n---------Reasoning Ended, Response Started---------")
    elif t == "content_streaming":
        print(chunk.get("content", ""), end="", flush=True)

Google (Gemini 2.5)

Python
JavaScript
TypeScript

async for chunk in await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="google",
    model_name="gemini-2.5-pro",
    thinking={"budget_tokens": 8192},
    stream=True
):
    t = chunk.get("type")
    if t == "reasoning_streaming":
        print(chunk.get("content", ""), end="", flush=True)
    elif t == "reasoning_ended":
        print("\n---------Reasoning Ended, Response Started---------")
    elif t == "content_streaming":
        print(chunk.get("content", ""), end="", flush=True)

Google (Gemini 3)

Python
JavaScript
TypeScript

async for chunk in await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="google",
    model_name="gemini-3.1-pro-preview",
    thinking={"effort": "high"},
    stream=True
):
    t = chunk.get("type")
    if t == "reasoning_streaming":
        print(chunk.get("content", ""), end="", flush=True)
    elif t == "reasoning_ended":
        print("\n---------Reasoning Ended, Response Started---------")
    elif t == "content_streaming":
        print(chunk.get("content", ""), end="", flush=True)

xAI

Python
JavaScript
TypeScript

async for chunk in await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="xai",
    model_name="grok-4.3",
    thinking={"effort": "high"},
    stream=True
):
    t = chunk.get("type")
    if t == "reasoning_streaming":
        print(chunk.get("content", ""), end="", flush=True)
    elif t == "reasoning_ended":
        print("\n---------Reasoning Ended, Response Started---------")
    elif t == "content_streaming":
        print(chunk.get("content", ""), end="", flush=True)

OpenRouter

Python
JavaScript
TypeScript

async for chunk in await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="openrouter",
    model_name="minimax/minimax-m1",
    thinking={"max_tokens": 4096},
    stream=True
):
    t = chunk.get("type")
    if t == "reasoning_streaming":
        print(chunk.get("content", ""), end="", flush=True)
    elif t == "reasoning_ended":
        print("\n---------Reasoning Ended, Response Started---------")
    elif t == "content_streaming":
        print(chunk.get("content", ""), end="", flush=True)

Cerebras

Python
JavaScript
TypeScript

async for chunk in await client.send_message(
    "A driver was going the wrong way down a one-way street. He passed several police officers, but they did not stop him. Why? Think step by step.",
    llm_provider="cerebras",
    model_name="z-ai/glm-4.7",
    thinking={},
    stream=True
):
    t = chunk.get("type")
    if t == "reasoning_streaming":
        print(chunk.get("content", ""), end="", flush=True)
    elif t == "reasoning_ended":
        print("\n---------Reasoning Ended, Response Started---------")
    elif t == "content_streaming":
        print(chunk.get("content", ""), end="", flush=True)

Check Model Support

models = await client.list_models(supports_thinking=True, model_type="llm")
for m in models:
    print(f"{m.provider}/{m.name}")

See Models for details.

​Fields

​Provider Support

​Provider Details

​Non-Streaming Examples

​OpenAI

Python

JavaScript

TypeScript

​Anthropic

Python

JavaScript

TypeScript

​AWS Bedrock

Python

JavaScript

TypeScript

​Google (Gemini 2.5)

Python

JavaScript

TypeScript

​Google (Gemini 3)

Python

JavaScript

TypeScript

​xAI

Python

JavaScript

TypeScript

​OpenRouter

Python

JavaScript

TypeScript

​Cerebras

Python

JavaScript

TypeScript

​Streaming Examples

​OpenAI

Python

JavaScript

TypeScript

​Anthropic

Python

JavaScript

TypeScript

​AWS Bedrock

Python

JavaScript

TypeScript

​Google (Gemini 2.5)

Python

JavaScript

TypeScript

​Google (Gemini 3)

Python

JavaScript

TypeScript

​xAI

Python

JavaScript

TypeScript

​OpenRouter

Python

JavaScript

TypeScript

​Cerebras

Python

JavaScript

TypeScript

​Check Model Support

​Related

On this page

Fields

Provider Support

Provider Details

Non-Streaming Examples

OpenAI

Anthropic

AWS Bedrock

Google (Gemini 2.5)

Google (Gemini 3)

xAI

OpenRouter

Cerebras

Streaming Examples

OpenAI

Anthropic

AWS Bedrock

Google (Gemini 2.5)

Google (Gemini 3)

xAI

OpenRouter

Cerebras

Check Model Support

Related