Gemini 4 Argon arrives in Devin: 60.5% on FrontierCode at 75% off

Gemini 4 Argon is now available in Devin Desktop and Devin CLI. Through October 27, 2026, it costs 75% less than list price. Google keeps a 50% launch discount in place until December 1, and Devin adds another 50% for the duration of its promotion. On the in-house FrontierCode 1.1 benchmark, the model scores 60.5%. That is above GPT-6.1 Sol (60.4%) and on par with GPT-6 Sol (60.7%), and it is the best result among Gemini models in the table.

What FrontierCode shows

FrontierCode 1.1 is a proprietary benchmark from the Windsurf/Devin team. It evaluates models on real engineering tasks using two criteria: solution quality and mergeability, meaning whether the changes can go into a production repository without being rewritten. The metric is practical, but the methodology and task set cannot be independently verified. Treat the numbers as vendor data.

Key results from the table:

• Gemini 4 Argon: 60.5%. The gain over Gemini 3.7 Flash (56.3%) is about four points.
• GPT-6.1 Sol: 60.4%, GPT-5.6 Sol: 60.6%, GPT-6 Sol: 60.7%. Argon sits in a tight cluster with this group, and the 0.3-point spread is within noise.
• Ranked higher: SWE-2 (62.5%), Claude Sonnet 5.5 (64.4%), GPT-6 Astra (64.5%) and Claude Opus 5.5 (65.3%).
• Ranked lower: Grok 4.7 (59.4%), Claude Haiku 5.5 (58.4%) and Kimi K3 (58.2%).

Among the 12 models in the table, Argon ranks seventh. It cannot be called the quality leader. Its pitch is the combination of price and performance, and on that front it currently holds a strong position.

How the model behaves at work

According to the Devin team, Argon gathers context first and writes code only afterward. It reads widely, checks git history to pick up the repository's conventions, and reviews its own diff before declaring a task done. This sequence directly affects mergeability: fewer edits that break the project's style and fewer overlooked spots.

The model works best on bug fixing and refactoring in backend and systems code, especially in Python, C/C++ and PHP. The publication says nothing about frontend, mobile development or other stacks, so it is too early to draw conclusions about them.

Pricing and how to test it

The discounts do not add up arithmetically. Google takes half off the price, and Devin cuts the remainder in half again, so the final price is a quarter of list. The publication does not give absolute rates. On the cost-versus-quality chart, the cost axis shows values per rollout from $0.25 to $10, and Argon appears in two variants: at regular price and with the promotion.

The practical takeaway for teams is simple. Before October 27, you can run Argon on your own tasks alongside your current default model: take 20–30 closed tickets covering bug fixes and refactoring, then measure the share of diffs accepted without edits and the cost per task. If you write in Python, C/C++ or PHP, the result will be the most telling. After October 27, Devin's discount ends and only Google's launch rate remains, until December 1. Before scaling up, it is better to build the non-promo price into your calculations to avoid an unpleasant budget surprise.

You can install the model in Devin Desktop or through Devin CLI.