Generation Parameters & Safety

The knobs that change every reply

Six parameters control most of what changes between two otherwise-identical Gemini calls:

temperature (0.0–2.0) — randomness. 0 is near-deterministic; 1.0 is the balanced default; > 1 is creative-but-weird.
topP (0.0–1.0) — nucleus sampling. Restrict to tokens making up p of the probability mass.
topK (1–100+) — only consider the top-K candidate tokens.
maxOutputTokens — hard cap. Hitting it triggers finishReason: MAX_TOKENS.
responseMimeType — text/plain by default; switch to application/json for JSON mode.
seed — pin the random seed for reproducible runs (best-effort, not a guarantee on Pro).

Thinking config — the new lever

Gemini 2.5 introduced explicit thinkingConfig. The model can spend extra compute thinking before answering, and you pay for those thinking tokens at output rates. Choose the level deliberately:

MINIMAL — no extended thinking. Cheapest, fastest, fine for chat and simple Q&A.
LOW — small budget. Useful for moderately tricky reasoning.
MEDIUM — default for hard problems.
HIGH — long-form reasoning, complex code, multi-step math. Spend this only when you mean to.

Safety settings

Four harm categories with five threshold levels each. Defaults in Gemini 2.5+ are OFF — the model does not block by default; you opt in to filtering.

Category	What it covers
`HARM_CATEGORY_HARASSMENT`	Targeted identity attacks
`HARM_CATEGORY_HATE_SPEECH`	Slurs, profanity directed at groups
`HARM_CATEGORY_SEXUALLY_EXPLICIT`	Sexual content
`HARM_CATEGORY_DANGEROUS_CONTENT`	Instructions for harmful acts

Thresholds: OFF, BLOCK_NONE, BLOCK_ONLY_HIGH, BLOCK_MEDIUM_AND_ABOVE, BLOCK_LOW_AND_ABOVE. The difference between OFF and BLOCK_NONE: OFF disables the classifier entirely; BLOCK_NONE still classifies but never blocks.

Code

Generation params in Python·python

from google import genai
from google.genai import types

client = genai.Client()

response = client.models.generate_content(
    model='gemini-2.5-flash',
    contents='Write a 3-sentence product description for a $40 espresso scale.',
    config=types.GenerateContentConfig(
        system_instruction='You are a professional copywriter.',
        temperature=0.4,
        top_p=0.95,
        top_k=40,
        max_output_tokens=300,
        seed=42,
        # Thinking budget — opt in for hard problems
        thinking_config=types.ThinkingConfig(thinking_level='LOW'),
    ),
)
print(response.text)

Tighten safety on a kid-facing app·python

config = types.GenerateContentConfig(
    safety_settings=[
        types.SafetySetting(
            category=types.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT,
            threshold=types.HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
        ),
        types.SafetySetting(
            category=types.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
            threshold=types.HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
        ),
        types.SafetySetting(
            category=types.HarmCategory.HARM_CATEGORY_HATE_SPEECH,
            threshold=types.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
        ),
    ],
)

Exercise

Run the same prompt three times: once with temperature 0.0 (and a fixed seed), once with 0.7, and once with 1.5. Compare outputs. Then enable thinking_config: MEDIUM on a moderately hard logic puzzle (something like "Three people split a $30 dinner; each pays $10. The waiter realizes it should be $25 and returns $5..."). Compare the thinking-on vs thinking-off answers. Write 2–3 sentences on what you noticed.