
It looks like Google has released a new batch of AI models. They go by the name of Argon, Carbon and Hydrogen, all under a new Gemini-4 model umbrella. Gemini Carbon is at least as good if not better than Anthropic’s Claude Fable 5.2, but has the advantage of being significantly cheaper.
Model Differences & Capabilities:
Gemini 4 Carbon (google/gemini-4-carbon)
- Role: Google’s strongest frontier model built for deep, long-horizon reasoning on the hardest problems.
- $2.00 input / $10.00 output (introductory price through November 30, 2026; rising to $4.00 input / $20.00 output on December 1, 2026).
Gemini 4 Argon (google/gemini-4-argon)
- Role: The balanced mid-tier model for general knowledge work and long-horizon tasks with lower token overhead than Carbon.
- $2.00 input / $10.00 output (introductory price through November 30, 2026; rising to $4.00 input / $20.00 output on December 1, 2026).
Gemini 4 Hydrogen (google/gemini-4-hydrogen)
- Role: A high-speed, high-efficiency execution model built for rapid coding and agentic workflows. Similar to previous Flash models.
- Gemini 4 Hydrogen is priced flat across its entire 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens.
With this latest drop, I decided to use Carbon (go intro pricing!) and the new Orchestrator controller, which can spawn several sub-agents, with each sub-agent using Gemini 4 Hydrogen, to achieve a goal. It seemed the easiest way to leverage sub-agents. However, Gemini comes with a plethora of commands to do this manually, i.e. /boost, /teamwork and /swarm.
I had Gemini convert an old Star Trek game that I used to play on a DEC VAX computer when I was in high school, from C to Rust. It was the epitome of geekiness at the time, and I loved it as much as the kids today love Grand Theft Auto.
After some simple prompts to get things going, Gemini spawned up to 6 sub-agents. One of them wrote a C to Rust converter and confirmed an exact duplication of code. Another agent found bugs, another agent fixed those bugs and so on. The entire process was done in about an hour. Gemini converted the front-end UI into WASM, something I did not even ask for explicitly.

One year ago, I was impressed with AI coding agents, but was discouraged by the hallucinations and bugs. Now, I am just impressed. While certainly not perfect, they are incredibly powerful tools that can amplify any software engineer/developer to a whole new level.
The code can be viewed here: https://github.com/nbrandaleone/trek

The unreasonable effectiveness of Gemini in code generation was originally published in Google Cloud – Community on Medium, where people are continuing the conversation by highlighting and responding to this story.
Source Credit: https://medium.com/google-cloud/the-unreasonable-effectiveness-of-gemini-in-code-generation-e30e486d97da?source=rss—-e52cf94d98af—4
