Think Less, Pay Less: Routing Agent Subworkloads by Thinking Level in Gemini 3.7 Flash

Gemini 3.7 Flash ships a per-request thinking_level parameter so your agent can dial reasoning up for hard steps and down for cheap lookups — one model, one endpoint, variable cost.

Most agentic pipelines treat every step identically: every tool call, every lookup, and every complex code analysis goes to the same max-reasoning inference tier. Gemini 3.7 Flash (GA today) breaks that pattern with a thinking_level parameter in generation_config that lets you tune per call.

Three levels:

Level	Best for	Cost profile
low	Quick lookups, file listing, simple tool calls	Lowest
medium	Default for most agentic steps, code analysis	~1.5× low
high	Hard debugging, complex code gen, architectural design	Highest

Routing pattern:

from google import genai

client = genai.Client()

def step(task: str, thinking_level: str) -> str:
    return client.interactions.create(
        model="gemini-3.7-flash",
        input=task,
        generation_config={"thinking_level": thinking_level},
    ).output_text

# Cheap: simple context-prep steps
file_list = step(
    "List all Python files modified in the last 24h",
    thinking_level="low",
)

# Balanced: most analytical steps
plan = step(
    f"Given these files: {file_list}\n"
    "Identify which ones might contain auth logic.",
    thinking_level="medium",
)

# Full depth: complex debugging
fix = step(
    f"Review this auth module: {plan}\n"
    "Find any race conditions and rewrite the transaction lock safely.",
    thinking_level="high",
)

Via OpenRouter (OpenAI-compatible client):

import openai

client = openai.OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="<openrouter-key>",
)

resp = client.chat.completions.create(
    model="google/gemini-3.7-flash",
    messages=[{"role": "user", "content": task}],
    extra_body={"generation_config": {"thinking_level": "high"}},
)

Why this matters for Claude Code workflows: On days when Anthropic quota is under pressure, Gemini 3.7 Flash is a solid background-task model — particularly for the compute-heavy "write a migration plan" or "find the bug" steps where you want real reasoning but don't need Claude's specific strengths. The thinking-level routing keeps those background costs proportional to actual task complexity.

Sources: Gemini 3.7 Flash — ai.google.dev · Gemini thinking guide · Gemini 3 Developer Guide