Gemini 4 Carbon is another reason to choose AI by the work.

I think this will start to feel a little like streaming. You might keep Netflix because you want to watch Stranger Things, while subscribing to HBO for Game of Thrones. You are choosing each service for something specific that you cannot get in quite the same way elsewhere. Work with AI may look similar. One team might prefer Gemini for a task that fits its Google environment, Claude for a long document or coding workflow, and ChatGPT for a different kind of research or collaboration. The point is to know why each tool has a place, rather than paying for every new model because it won a headline.

Google's own comparison table explains why this choice cannot come from one overall score. Carbon leads in some evaluations, while Claude Opus 5.5 scores higher on an OSWorld computer-use test and GPT-6 Astra scores higher on FrontierSWE 2. Artificial Analysis found Carbon answered 58% of questions correctly on AA-Omniscience v1.1, with an 11% hallucination rate. That rate rose from 4% for Gemini 4 Argon, yet remained lower than the rates for the compared Claude and GPT models on this particular test. It does not tell us how often any of these models will make an error in a client's Gmail or Drive workflow.

That is why the carousel of models needs a purpose. A customer-support agent reading email has different risks from a tool helping someone brainstorm a presentation. A quarterly investor package needs its numbers checked against the source spreadsheets. A research assistant needs to show where it found a claim. If several models are involved, employees should know which tool is approved for each task, what information it can reach, who reviews the result, and what the work actually costs. Otherwise, a company can end up with more subscriptions without a clearer way of working.

Carbon is not yet something most teams can put into a Google Workspace workflow. Google says its first access is going to a subset of cyber defenders and then other trusted partners, with broader access to follow. When it reaches business users, I would start with a task the team already knows and compare Carbon with the model they use today. Did it find the right information? How much did a person need to correct? Was the result worth the time and cost? A real workflow tells us more than a model launch can.

At Ruby, we are not tied to one model. We learn how a client's team actually works, map the repeated questions and handoffs, and test the tools against those processes. Then we help employees learn the workflow that fits, including where to check a source and when to take over. Our ongoing support matters here too: if next month's model handles a task better, we can bring that change into the work instead of asking the team to start from scratch. The goal is not to build the biggest carousel. It is to give people the right set of tools and the confidence to use them.