Google tests Carbon-H Gemini 4 build as Argon 4 on Arena
Testers say the Carbon-H checkpoint is faster at coding, design and 3D work, though front-end development stays a weak spot.
Google is testing a new Gemini 4 Carbon checkpoint under the name Argon 4 on the Arena model evaluation platform, with the build labeled Carbon-H, according to aitimes.com. Testers report that Carbon-H is faster and stronger at coding, design and 3D modeling than earlier Carbon builds.
Some tests show the model spending 8 to 10 minutes reasoning before it produces output. Web development and front-end UI work remain weak points, according to the testers.
Mixed results in head-to-head comparisons
In one comparison run by X user OltAi, a pixel-style T-Rex animation was generated across several models. Carbon scored 9.5 out of 10, followed by Claude Opus 5.5 at 9.3, GPT-6 Astra at 8.8 and GPT-6.1 Terra at 4.9.
Google plans to price Carbon at 2 dollars per million input tokens and 10 dollars per million output tokens, the same as Argon's initial launch price. The company has not issued an official statement on the update, and Arena does not clearly disclose which checkpoint was used in its tests.
On October 28, 2026, Google released Gemini 4 Carbon for cybersecurity professionals, with access limited to some security experts. In earlier tests of that first Carbon version, 3D work drew praise while processing was slow and front-end coding drew criticism.
Quick answers
What is Carbon-H?
Carbon-H is the label on a new Gemini 4 Carbon checkpoint that Google is testing on Arena under the name Argon 4.
How much will Gemini 4 Carbon cost?
Google plans to price it at 2 dollars per million input tokens and 10 dollars per million output tokens, the same as Argon's initial launch price.
What are Carbon-H's weak points?
Testers say web development and front-end UI work remain weak, and some tests take 8 to 10 minutes of reasoning before output.