What Gemini 4 Argon Is Actually Good For

Gemini 4 Argon is best understood as a specialist within a model stack, not a replacement for every other model. It became available for testing two weeks after its announcement and is described as a new-pretrain, ultra-large model in the same class as Astra and Fable; Google has not disclosed its size. Its strongest case is knowledge work: it leads the cited VALS index, Automation Bench, VALS Finance Agent, and Harvey’s legal-agent benchmark, where it scored 13 points ahead of Fable 5.1. Google also reported state-of-the-art performance in agentic science, cybersecurity, and multimodal tasks, but the hosts caution that the picture depends on which benchmarks Google chose to highlight.

Argon’s coding performance is uneven rather than uniformly frontier-leading. It leads DeepSwee by a few points over Opus 55, but trails Astra by about 10 points on FrontierSwee and Opus 55 by 9 points on Terminal Bench 4.0. It scored 69.2% on OSworld, just below Astra’s 72.6%. Artificial Analysis gave it a score of 53, tied for fourth with Fable 51 and GPT-6 Astra; Opus 55 led at 58. These results suggest that Argon’s usefulness depends on the coding task and evaluation, rather than establishing it as a clear top choice across software work.

Early user reports reinforce that variability, with the harness appearing to matter. Pavel Hurin first found Argon weak at bug hunting, then switched from Gemini CLI to the anti-gravity CLI and reported a frontier result close to Opus 5.5 Max, repeated over three runs. Max Weinbach likewise found it weak in anti-gravity but good in Claude Code and Open Code. However, the transcript notes that using Argon in other harnesses technically violates its terms of service. Other assessments were harsher: Noah Denigin called it useless for real-world code, while Dan Shipper reported a 21% score on EveryWrites’ Senior Engineer benchmark, below Soul 6 and Muse.

Shipper’s team found a more specific coding niche: Argon was fast and effective for greenfield apps, but shaky at refactoring and prone to skipping self-verification in existing codebases. After six Every staff members tested it for a week, Shipper recommended it as a strong option for companies in Google’s ecosystem, and as a secondary model for editorial collaboration, quick prototypes, and steerable UI or dashboard design. He preferred using it through Cursor or the API. For his own writing, he said it became his favorite partner for structuring difficult essays, sharpening metaphors, and close-reading books.

Multimodal work may be another area of strength, consistent with Google’s history in that area, though the demonstrations vary in scope and cost. Examples included BattleChess graphics described as AAA-quality, a walkthrough of a large pagoda complex, and a 3D Apple-computer museum built from archival images and specifications. The pagoda was produced through a costly multi-agent run, not a one-shot generation, so it should not be treated as evidence that the same result is effortless or inexpensive.

The bugged dev used a single image of a Gothic Victorian carriage to create a 174,000-face interactive 3D model, with opening doors, hinged windows, wheel physics, suspension, and flickering lanterns, using code rather than Blender. A procedurally generated 3D battle helicopter was another early example. But creative results were not consistently compelling: Max Techno found Argon less impressive than Opus 5.5 or Astra on some tasks, citing washed-out colors, floating textures, and dated interface design. Argon’s API price is reduced by 50% through November 30, and its Artificial Analysis benchmark run cost $199 per task during the discount; afterward, it is expected to cost more than Astra but about two-thirds as much as Anthropic’s models. The hosts characterize it as a discounted frontier model whose appeal depends on the work, workflow, and price.