thinking_level Guide

Gemini 3.5 Flash replaced the numeric thinking_budget with four named levels. Understanding the difference is critical because the default changed with the model upgrade.

⚠️

The default changed — this is the silent trap

gemini-3-flash-preview defaulted to high (~8K thinking tokens). gemini-3.5-flash defaults to medium (~4K thinking tokens). If you don't set thinking_level explicitly, you get 50% fewer thinking tokens with no error or warning.

Level comparison

Level	Thinking Tokens	Typical Latency	Relative Cost	Best For
minimal	~1K	~1–2s	1×	Fact retrievalShort Q&AClassification
low	~2K	~2–4s	2×	Conversational chatSimple summarizationIntent detection
mediumnew default	~4K	~4–8s	4×	Code generationDocument summarizationRAG synthesis
highold default	~8K	~8–15s	8×	Code agentsMulti-step reasoningComplex code review

Config snippets per level

minimal~1K thinking tokens · ~1–2s · cost 1×

Best for high-volume, low-complexity tasks where latency and cost dominate.

✓ Good for

· Fact retrieval
· Short Q&A
· Classification
· Entity extraction

✗ Not ideal for

· Multi-step reasoning
· Code review
· Long-form writing

Python

generation_config=GenerationConfig(
    thinking_config=ThinkingConfig(
        thinking_level="minimal",  # ~1K thinking tokens
    )
)

TypeScript

generationConfig: {
  thinkingConfig: {
    thinkingLevel: "minimal",  // ~1K thinking tokens
  },
}

low~2K thinking tokens · ~2–4s · cost 2×

Good default for chat assistants that don't need deep reasoning.

✓ Good for

· Conversational chat
· Simple summarization
· Intent detection
· Slot filling

✗ Not ideal for

· Code agents
· Complex analysis
· Multi-step planning

Python

generation_config=GenerationConfig(
    thinking_config=ThinkingConfig(
        thinking_level="low",   # ~2K thinking tokens
    )
)

TypeScript

generationConfig: {
  thinkingConfig: {
    thinkingLevel: "low",   // ~2K thinking tokens
  },
}

medium~4K thinking tokens · ~4–8s · cost 4×⚠ new default — verify this is right for your use case

NEW DEFAULT for gemini-3.5-flash. Was "high" in gemini-3-flash-preview.

✓ Good for

· Code generation
· Document summarization
· RAG synthesis
· Moderate analysis

✗ Not ideal for

· Complex multi-step agents
· Deep code review
· Research synthesis

Python

generation_config=GenerationConfig(
    thinking_config=ThinkingConfig(
        thinking_level="medium",  # ~4K thinking tokens (NEW DEFAULT)
    )
)

TypeScript

generationConfig: {
  thinkingConfig: {
    thinkingLevel: "medium",  // ~4K thinking tokens (NEW DEFAULT)
  },
}

high~8K thinking tokens · ~8–15s · cost 8×← was default in gemini-3-flash-preview

OLD DEFAULT for gemini-3-flash-preview. Use for tasks that were working well before migration.

✓ Good for

· Code agents
· Multi-step reasoning
· Complex code review
· Research synthesis
· Agentic loops

✗ Not ideal for

· Simple Q&A
· High-volume batch
· Real-time chat

Python

generation_config=GenerationConfig(
    thinking_config=ThinkingConfig(
        thinking_level="high",   # ~8K thinking tokens (OLD DEFAULT)
    )
)

TypeScript

generationConfig: {
  thinkingConfig: {
    thinkingLevel: "high",   // ~8K thinking tokens (old default)
  },
}

Quick decision guide

You were using gemini-3-flash-preview and it worked well

thinking_level="high" — restore previous reasoning quality

High-volume chat assistant (cost matters)

thinking_level="low" — halves cost vs new default

RAG pipeline with document synthesis

thinking_level="medium" — new default is fine here

Code agent or agentic loop

thinking_level="high" — complex multi-step tasks need it

Classification or entity extraction at scale

thinking_level="minimal" — minimum cost for simple tasks