Gemini last models: temperature, top_p, and top_k are deprecated and ignored

TL;DR Highlight

Gemini 4 Argon ignores temperature, top_p and top_k; use thinking_level instead

Who Should Read

A backend developer who has tuned sampling parameters in the Gemini API and is migrating to Gemini 4 Argon

Core Mechanics

Sampling parameters are ignored
- temperature, top_p and top_k are ignored if you send them.
- The migration checklist says to remove all three from your requests.
thinking_level controls depth
- There are three levels: low, medium (default) and high.
- It is the main control over quality, latency and token use.
Output limits got much bigger
- Input, thinking and output share one 1,048,576-token window.
- With background execution a single response can go past 64K, up to 1M tokens.
Raise max_output_tokens
- The docs say Gemini 4 Argon can spend more of the limit on thinking than earlier models.
- If your limit is under 32K, raise it to at least 32,768.
Explicit caching is gone
- cachedContents / caches.create is not supported.

Community Reactions

Doubts about prompt-only determinism
- One commenter questioned whether a system instruction is guaranteed to work better than top_k or top_p.
- Another said it amounts to a mini Agent.md.
Pain for diversity-heavy workloads
- A commenter said parallel research sub-agents need diverse hypotheses, otherwise running five equals running one.
- They said the latest OpenAI models did the same, so they add random variation to prompts to compensate.
Reasoning models explanation
- A commenter said reasoning models are trained around specific configurations, so sampling is replaced by reasoning effort.
- Another guessed that RL training makes models brittle to parameter changes.
Other providers are doing it too
- One comment said sonnet-5 also drops temperature.
- Another said these parameters have been nearly useless in recent model generations and often made results worse.

How to Apply

Remove sampling params from requests
- Search your codebase today for temperature, top_p and top_k in gemini-4-argon calls and delete them.
- Clean them up now since later generations return HTTP 400.
Run evals at all three thinking levels
- Run your existing eval set at low, medium and high.
- Pick the lowest level that meets your quality bar.
Raise max_output_tokens and timeouts
- Raise limits under 32K to at least 32,768.
- Use background execution or set HTTP timeouts to 15+ minutes for hard prompts.
Drop explicit cache calls
- Remove calls that create or reference cached_content.
- Put the shared prefix first so prompts of 4,096+ tokens get implicit caching discounts.

Try It

copy & run

# Paste into your Gemini API code and run in Python (pip install google-genai, set GEMINI_API_KEY)

from google import genai
client = genai.Client()

for level in ["low", "medium", "high"]:
    interaction = client.interactions.create(
        model="gemini-4-argon",
        input="{{질문}}",
        generation_config={"thinking_level": level, "max_output_tokens": 32768},
    )
    print(level, interaction.usage)
    print(interaction.output_text[:300])

# Next: replace {{질문}} with a real eval question and compare quality and token use per level.

Terminology

temperature / top_p / top_kDials that control randomness when picking the next token. Higher values give more varied answers, lower values more similar ones.

thinking_levelA low, medium or high setting for how deeply the model reasons before answering. Higher means more tokens and a slower first token.

long decodeGenerating a response in 64K-token hops chained up to 1M tokens. Like writing a long piece in continued chunks.

Implicit cachingThe server automatically reuses repeated prompt prefixes without setup, cutting input cost.

background executionAn async mode where the server runs your request and you poll for the result.