Gemini 4 Argon

TL;DR Highlight

Gemini 4 Argon ties #1 on CWE-Bench (68%) and sped up Gemma 4 E2B on vLLM 2.85x

Who Should Read

A backend developer choosing a new LLM for coding agents or large-scale refactors

Core Mechanics

Argon ties for #1 on CWE-Bench
- It scores 68% on CWE-Bench for vulnerability remediation, tied for first.
- Google calls it its most resilient model against indirect prompt injection.
Chrome refactor cut to weeks
- It helped separate a ~6M-line legacy browser-features folder inside Chromium's 40M-line codebase.
- The refactor spanned hundreds of files and took weeks instead of months, per the post.
Gemma 4 E2B inference got 2.85x faster
- On an A100 via vLLM, throughput rose from ~37,900 to over 108,000 tokens per second.
- The post reports 100% token parity, no quality drop, and a few days of work.
ADK optimized overnight
- The team gave Argon their .agents/skills rulebook and a one-paragraph goal.
- In-memory session prep time dropped 40% and cold imports got 21% faster, now live in google/adk-python.
xTB and lab migration also improved

Community Reactions

Commenter reports reasoning-loop bugs
- A tester using 100 coding environments placed it between Fable 5 and 5.1.
- They saw infinite reasoning loops and broken reasoning blocks, giving a higher failure rate than rivals.
Commenter found it costly and slow
- On 11 small bug fixes it solved 10, with median cost $0.52 vs $0.18 for Sonnet and $0.09 for Sol.
- Runs took ~4 minutes vs 32 seconds and ~1.5 minutes, with ~24k output tokens vs ~3k.
Commenters praised writing and non-Western context
- One said answers are concise, logical, and distinctively written.
- Another found it accurate on Sri Lankan and Tamil cultural topics.
Many complained about the rollout
- Some said it is Ultra-only, leaving Pro users without it.
- One reported the Gemini 4 option vanished after paying.

How to Apply

Rerun one agent task on Argon today
- Take 1-3 bug fixes you already run on your current model and run the same prompts via OpenRouter or AI Studio.
- Record success, time, output tokens, and cost side by side.
Set token caps and timeouts
- Comments reported infinite loops and a 10-minute timeout, so cap max output tokens and wall-clock time.
- Add a fallback that retries with another model on timeout.
Try an overnight rulebook-plus-goal run
- Keep a rules file like .agents/skills in the repo and state must-not-break constraints in a one-paragraph goal.
- Run a perf or migration task overnight and review the results as separate PRs.

Try It

copy & run

# Paste into Claude, ChatGPT, or Claude Code (simplified version of the ADK case in the post)

You are a performance engineer on this repository. Read the rulebook below first and follow it.
[Rulebook]
{{.agents/skills or your coding rules}}

[Goal]
{{Reduce framework overhead without breaking existing apps}}

Workflow:
1. Decide how to profile and which metrics to use.
2. Find 3 bottleneck candidates and give evidence.
3. Split changes into small independent PRs.
4. For each change, report test results and before/after numbers.

Next step: run the same prompt on two models and compare number of changes, test pass rate, and cost.

Terminology

CWE-BenchA benchmark measuring how well a model fixes categorized software vulnerabilities (CWE). Think of it as an exam for security patching.

Indirect prompt injectionAn attack where instructions hidden in external data like web pages or documents trick the model. Like a fake note slipped inside a delivered package.

AntigravityGoogle's agentic development platform. In the post, it is where Argon does long autonomous work.

Blender MCPA connector that lets agents operate Blender as a tool. Used in the post to build 3D scenes.

Speculative decodingA small model drafts tokens and a larger model verifies them to speed up generation. Like writing a draft and getting it reviewed.