Google Opens Gemini 4 Carbon: An Agent-Focused Model at $2/$10 Until December

Google has rolled out Gemini 4 Carbon, and it is not yet another "general-purpose model" but a tool aimed at a specific pain point. It is about an agent that has to poke around a codebase for a long time or methodically test hypotheses, rather than give an answer on the first try. Carbon is built on top of Argon but is tuned to think longer.

How Carbon Differs from Argon

In short: same foundation, more reasoning. Google says outright that Carbon spends more tokens on "thinking" than Argon. For tasks where a mistake is costly, that is a reasonable trade. For example, finding the cause of a bug in someone else's repository or working out why a pipeline is failing. You pay in time and tokens, and in return you get a chance at a more careful result.

Google's demo is impressive. The model built an interactive 3D simulator that synthesizes the songs of six bird species on the fly. You can open a bird's beak and listen to how the sound changes. But a single demo is not a metric. It is best read as an example of a long, multi-part task in which graphics, sound synthesis, and interactivity came together in one project.

What the MKA Benchmark Showed

The new MKA test is closer to office reality than to olympiad-style problems. The agent is given a batch of engineering drawings and other files from which it must assemble a report, a spreadsheet, or a presentation. The test checks not whether the model can write a function, but whether it can make sense of heterogeneous input and produce a finished deliverable.

On the hard criteria, Carbon High scores 54%, while Fable 5.2 Max scores 57%. Carbon cannot be called the leader here: it trails its competitor slightly. Three percentage points on a single benchmark is not a verdict, but it is also no reason to believe headlines about the "best model for agents." The test is also new, and how it relates to your real data is not yet clear. The benchmark page is hosted on Artificial Analysis.

Access and Pricing

The model is already available in the paid API. In the Gemini chat and in Antigravity, only Ultra subscribers get it.

API prices are as follows:

• through the end of November: $2 per million input tokens and $10 per million output tokens;
• from December 1: $4 and $20, that is, twice as expensive.

There is a catch. Since Carbon spends more tokens on reasoning, the real cost should be calculated not from the per-million price but from the total tokens per solved task. A low per-token price for a "wordy" model does not guarantee a cheap bill.

What to Do in Practice

Don't trust either the marketing or any single benchmark. It is better to run a simple experiment.

1. Take a set of your own real agentic tasks: bug analysis, code review, assembling a report from a set of files.
2. Run it on Argon and on Carbon and measure two things: the share of tasks solved and token consumption per task.
3. Do this before December 1, while the $2/$10 price applies. Then the decision to switch will be based on numbers.

Product teams should build the December price of $4/$20 into their unit economics right away, not the introductory one. If Carbon beats Argon only on rare hard cases, it may make sense to enable it selectively, as "heavy artillery" for difficult tasks rather than as the default model.

Limitations

All figures here are taken from Google's announcement and the benchmark page. The source contains no independent verification. The publication also gives no comparison with Argon on quality and token consumption, so you will have to measure that yourself. Google has put the model's details on a separate card.

Bottom line: Carbon is a sensible step toward "thinking" agents, but not an obvious winner. The 3-percentage-point gap behind Fable 5.2 Max on MKA and the price doubling from December make testing on your own tasks a must before adoption.