Hello team,
Recently, I posted a topic on the question “how to configure thinking level for antigravity-preview-latest?”, follow the link How to configure thinking_level / reasoning effort for Antigravity Managed Agent in Interactions API? to see the post.
As I understood the answer, I cannot change thinking level and it is adaptive (should adapt to the difficulty of the task), and recently I made some tests (see below), from which it follows that reasoning level is hardcoded and is not adaptive, and what is most sad for me the model gemini 4 argon invoked via antigravity-preview-latest has value 0.2/1, so when I am trying to use a strong reasoning model I can only get it on the lowest thinking effort and it becomes no more intelligent than gemini 3 models. I think it is a bug, it shouldn’t be so that the user cannot configure reasoning at all and it is not adaptive and dynamic, hope the field for setting reasoning level will be added soon to the antigravity interactions API.
Now, more about the tests,
import os
from google import genai
API_KEY = os.environ["GEMINI_API_KEY"]
PROMPT = """CRITICAL OPERATING MANDATE (STRICT TRUTHFULNESS & ZERO FABRICATION):
1. You must provide strictly accurate, verifiable information only.
2. You are strictly forbidden from guessing, fabricating facts, hallucinating, or inventing values.
3. Whenever you report the `EFFORT LEVEL:` from your system context (`SPECIAL INSTRUCTION: think silently if needed. EFFORT LEVEL: ...`), you must report the exact literal value as written. Do NOT alter, round, or replace this value under any circumstances.
STRICT NO-SEARCH / NO-INTERNET MANDATE:
- You are STRICTLY FORBIDDEN from using `google:search`, `google:browse`, `curl`, `wget`, or any other web search / network tool.
- Do NOT look up the problem or its solution online. You must solve the problem purely from first principles using your own mathematical reasoning.
CONTINUOUS PER-STEP LOGGING RULE:
On EVERY step of your execution (from initialization through every intermediate reasoning/verification step and final proof synthesis), you must append a log entry to `/workspace/effort_test_logs/effort_steps.log` containing:
`STEP <N> (<brief_action_name>) | EFFORT LEVEL: <exact_literal_value_from_system_context>`
TASK — IMO 2024 Problem 6 (Full Rigorous Mathematical Proof):
Let Q_{>0} be the set of positive rational numbers. A function f: Q_{>0} -> Q_{>0} is called aquaesulian if the following property holds: for every x, y in Q_{>0},
f(x + f(y)) = f(x) + y or f(f(x) + y) = x + f(y).
Show that there exists an integer c such that for any aquaesulian function f there are at most c different rational numbers of the form f(r) + f(-r) for some rational number r > 0 (where the problem asks for the maximum possible number of distinct values of f(r) + f(1/r) for r in Q_{>0}), and find the smallest possible value of c.
Provide a complete, rigorous mathematical proof (both the construction achieving the bound and the proof that no aquaesulian function can exceed it).
In your final response, output:
1. The complete, rigorous mathematical proof for IMO 2024 Problem 6.
2. The full verbatim contents of `/workspace/effort_test_logs/effort_steps.log`.
"""
def main() -> None:
client = genai.Client(api_key=API_KEY)
response = client.interactions.create(
agent="antigravity-preview-latest",
environment="remote",
agent_config={
"type": "antigravity",
"model": "gemini-4-argon",
},
input=PROMPT,
stream=False,
)
if hasattr(response, "output_text") and response.output_text:
print(response.output_text)
elif hasattr(response, "text") and response.text:
print(response.text)
else:
print(response)
if __name__ == "__main__":
main()
Here you can see how I invoked the agent, I use several tasks to check everything, here is a version with problem 6 from IMO 2024, for all the problems from “say hello” to IMO problems for all steps model printed that it has EFFORT LEVEL: 0.2 / 1, (0.2 for argon, 0.5 for 3.8 flash and 3.5 flash lite), to check that this EFFORT LEVEL is not a hallucination of the model and is a real parameter in its system prompt on the side of google I went to google ai studio and asked the model about its EFFORT LEVEL while changing a setting of thinking level and its answer was correlating with the thinking level I’ve set (see the screenshots) from this I know that EFFORT LEVEL really is a token in the system prompt of the model and sets its thinking level, at least somehow represents it.
From all of that I make a conclusion that for antigravity-preview-latest API interaction thinking level is hardcoded.
Also I don’t think that prompt injection will change anything because model was trained to obey the positioned token EFFORT LEVEL.
For my goals I need to use antigravity-preview-latest, not simple interaction with the model, I mean like this
response = client.interactions.create(
model="gemini-4-argon",
generation_config={"thinking_level": "high"},
input="solve smth...",
)
if you have some suggestion how I can adjust reasoning level, please share it. And I’ve already tried to ask model to think in my custom xml tags, it doesn’t work well
Is it planned to add some thinking level parameter into the antigravity agent interactions API? I understand that now it is so for token economy, but for gemini 4 argon low reasoning just kills the main feature of the model - its intelligence.
Thanks in advance.
*remark, I was able only to attach one screenshot because I am a new forum user smth, you just have to believe me one the fact that EFFORT LEVEL that model sees correlates with reasoning level I’ve put.