Two Hours of Carbon: Gemini 4 Carbon vs Argon on Real Code

Google made Gemini 4 Carbon generally available on November 10. I started my bench at 9:04 the next morning, four sessions in parallel, on the $99.99 Google AI Ultra plan. At 11:08 every session stopped:

Individual quota reached. Please upgrade your subscription to increase your limits. Resets in 2h5m5s.

— Antigravity CLI, two hours in

Thirteen runs had finished. Four died mid-fix. This was medium effort, not high.

Same Price, More Tokens

Carbon is the follow-up to Argon, which I tested a month ago. The model card says it is “based on Gemini 4 Argon”. The price is identical:

So what do you get for the same money? Google’s answer is in the launch post: Carbon works by “using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops.” Same price per token. More tokens.

The Same Six Changes

Same bench as the Argon and Fable 5.2 posts: six merged changes from a production monorepo, each replayed on a clean snapshot, graded by the tests the original fix shipped with. Same prompts, medium effort, 20 runs.

• Best score yet. 17 clean runs. No run broke a test that passed before.
• The lead is one run. Argon failed one of three runs on Change D. Carbon passed all three. Every other change came out the same.
• Change C still beats everyone. Carbon failed it three times, on the same four tests, like every model before it.
• It is slow. A fix took 30 minutes on average. The longest took 46.
• It writes the most. 1.82M output tokens: a third more than Argon and four times Fable.
• Half-price tokens, full-price bill. Carbon costs a quarter of Fable per token and landed on the same total. From December 1 the same 20 runs cost about $121.

What $99.99 Buys

The 17 runs before the quota stop add up to 8.3 hours of agent time and about $50 of tokens at list price. That is the window. It came back two hours later, and the last seven runs fit in the next one.

I ran four sessions at once, which is how Google wants you to use it. Antigravity shipped /swarm in the same release as Carbon: one coordinator, many parallel workers. Its own menu labels it “Preview & Quota Heavy”. They are not joking.

I am not the only one who hit the wall. kevmo314 on Hacker News:

Yeah, I was barely able to complete a single project with Ultra 5x using Carbon…

— kevmo314, Hacker News

One session at a time would probably last the window. I did not test that. At 30 minutes a fix, it would also deliver about ten fixes in five hours.

What Carbon Doesn’t Solve

• Speed. Artificial Analysis measured 67 seconds to the first token and 26 minutes per task, twice the median. majdalsado on Hacker News reported the same wait and called it “completely unusable at the moment”.
• Cost per fix. Carbon cost more than Argon for one extra clean run. Opus 5.5 did 13 clean runs for $38.
• The margin. One run on one change, in a different harness. Call it a tie with Argon and I won’t argue.
• Sample size. Six changes from one codebase. Another codebase could rank these models differently.

Lessons

• Carbon is Argon with a bigger budget. Same family, same price list, same fixes, more tokens.
• Check the plan before the model. The best score on my bench is no use if the quota stops the work at 11am.
• Parallel agents are sold faster than they are funded. Four workers emptied a $99.99 window in two hours.
• Run your own tasks before December 1. That is the cheapest Carbon will be.