Parameter Changes

Complete reference of every parameter that changed, was removed, or added new requirements in Gemini 3.5 Flash.

Change summary

Parameter	Status	Action Required
`temperature`	removed	Delete the parameter. Use thinking_level to control reasoning depth instead.
`top_p`	removed	Delete the parameter. No direct replacement.
`top_k`	removed	Delete the parameter. No direct replacement.
`thinking_budget`	deprecated	Replace with thinking_level="minimal\|low\|medium\|high". Remove thinking_budget.
`thoughts_in_content (thought preservation)`	default changed	Set thoughts_in_content=False for chat sessions, or truncate history explicitly.
`FunctionResponse.id`	now required	Add id field matching the function call id to all FunctionResponse objects.
`FunctionResponse.name`	now required	Add name field matching the function name to all FunctionResponse objects.

temperature / top_p / top_k removed

These sampling parameters are no longer accepted. They were used in gemini-3-flash-preview to control output diversity. In Gemini 3.5 Flash, output quality is primarily controlled via thinking_level.

Before (gemini-3-flash-preview)

# gemini-3-flash-preview — these params worked
from google.generativeai.types import GenerationConfig

generation_config = GenerationConfig(
    temperature=0.7,     # ← REMOVED in 3.5-flash
    top_p=0.9,           # ← REMOVED in 3.5-flash
    top_k=40,            # ← REMOVED in 3.5-flash
    max_output_tokens=2048,
)

After (gemini-3.5-flash)

# gemini-3.5-flash — remove temperature/top_p/top_k
from google.generativeai.types import GenerationConfig, ThinkingConfig

generation_config = GenerationConfig(
    # temperature, top_p, top_k removed — not supported
    max_output_tokens=2048,
    thinking_config=ThinkingConfig(
        thinking_level="high",  # use thinking_level for quality control
    ),
)

Thought preservation default changed

Watch out in multi-turn chats: Thinking tokens now accumulate in conversation history by default. A 10-turn chat at thinking_level="high" (~8K tokens/turn) can consume 80K+ tokens of context, increasing cost and approaching context limits.

Before — no thought accumulation

# gemini-3-flash-preview — thoughts NOT preserved by default
# Multi-turn chat: each turn only sees the output, not thinking tokens
chat = model.start_chat(history=[])
response1 = chat.send_message("Analyze this codebase...")
# response1.candidates[0].content has output only
response2 = chat.send_message("Now summarize your findings...")
# history is compact — no token blowup

After — thoughts accumulate (+ how to fix)

# gemini-3.5-flash — thoughts ARE preserved by default
# ⚠ Thinking tokens accumulate in chat history!
# A 10-turn chat with "high" level can exceed 80K tokens quickly

# Option A: Disable thought preservation for chat sessions
from google.generativeai.types import GenerationConfig, ThinkingConfig

generation_config = GenerationConfig(
    thinking_config=ThinkingConfig(
        thinking_level="high",
        thoughts_in_content=False,  # ← prevents accumulation
    ),
)

# Option B: Truncate history explicitly after each turn
def truncate_history(history, max_turns=5):
    return history[-max_turns * 2:] if len(history) > max_turns * 2 else history

FunctionResponse: id + name now required

This causes a 400 error — not a silent regression. Any FunctionResponse that omits id or name will fail with HTTP 400. Check all tool-use handlers.

Before — id and name optional

# gemini-3-flash-preview — id and name were optional
tool_response = {
    "functionResponse": {
        "response": {
            "result": "The weather in NYC is 72°F and sunny."
        }
    }
}
# This worked fine in gemini-3-flash-preview

After — both required

# gemini-3.5-flash — id AND name are REQUIRED
# Missing either field returns HTTP 400

tool_response = {
    "functionResponse": {
        "id": "call_abc123",      # ← REQUIRED: matches the function call id
        "name": "get_weather",    # ← REQUIRED: the function name
        "response": {
            "result": "The weather in NYC is 72°F and sunny."
        }
    }
}

# Python SDK equivalent
from google.generativeai.types import FunctionResponse

func_response = FunctionResponse(
    id=function_call.id,         # match the call's id
    name=function_call.name,     # match the function name
    response={"result": "72°F and sunny"},
)

← Full migration checklist

All steps in one place with copy-paste configs

Function calling gotchas →

400 errors, id+name requirement, thinking_budget conflict