Google’s latest large language model, Gemini 4 Argon, is now live in the Devin Desktop app and the Devin command‑line interface. For a short window it is offered at 75% off the regular list price. The discount combines Google’s standard 50% introductory rate (good through Dec 1 2026) with an extra 50% off that Devin gives until Oct 27 2026.
How the model performs
Devin tested Gemini 4 Argon on its own FrontierCode 1.1 benchmark, which measures how well models handle real‑world engineering tasks such as bug fixes and refactoring. Gemini 4 Argon scored 60.5%, edging out GPT‑6.1 Sol (60.4%) and matching GPT‑6 Sol (60.7%). It is the highest‑scoring Gemini model, four points ahead of the earlier Gemini 3.7 Flash. The blog notes that the model reads widely, checks git history for repository conventions, and reviews its own code diff before finishing, making it strongest on backend and systems code in Python, C/C++, and PHP.
Why it matters
Developers who use Devin can try a high‑scoring code‑assistant model at a steep discount, making it cheaper to experiment with AI‑driven code fixes. The benchmark results suggest Gemini 4 Argon may handle backend refactoring tasks better than some older models, though independent testing is still needed.
