Gemini 4 Argon scores 77.9% on DeepSWE v1.1: the full comparison deserves a closer look
Google presents Argon as a model for complex, multi-step tasks. It leads DeepSWE v1.1, but the order changes on Terminal-bench 4.0. Here’s how I’d evaluate it in a real development workflow.
Pool DezaFounder of IA Systems & Data1 min read
Google presents Argon as a model geared toward complex tasks that require sustaining reasoning across multiple steps, including software engineering (see the announcement).
For DeepSWE v1.1, its published table shows:
- Gemini 4 Argon: 77.9%
- Claude Opus 5.5: 74.2%
- GPT-6 Astra: 74.1%
Argon leads this comparison, 3.7 percentage points ahead of Opus.
But the order changes on Terminal-bench 4.0:
- Claude Opus 5.5: 66.4%
- GPT-6 Astra: 58.2%
- Gemini 4 Argon: 57.4%
Both results appear in the same table.
To evaluate Argon in a development workflow, I’d measure three things on real tasks from the repository:
- Changes that pass tests and review.
- Time it takes a person to fix them.
- Cost per accepted task.
Benchmarks help shortlist candidates. Testing against our own code is what lets us decide where to use them.
What result would make you switch the model you use for development?
