Gemini 4 Carbon is Google's best model. Google still says use Argon

You know the drill. A new frontier model drops, and by the next morning someone on the team wants to switch everything over. This time Google pulled the handbrake itself. The Gemini 4 Carbon launch post says it plainly: for most software engineering and knowledge work, Gemini 4 Argon remains the better fit. A vendor talking you out of its own flagship. You don't see that often.

What's new

• The model: Google calls Carbon its strongest model, built for cybersecurity, long-running agentic workflows, complex knowledge work, software engineering and scientific reasoning. It spends more tokens than Argon because it reasons longer and verifies more.
• Access: paid API via Google AI Studio and Gemini for business; rolling out to AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity. According to heise, only selected security experts had access for the two weeks before.
• Price: $2 per million input tokens and $10 per million output tokens through November 30, then double from December 1: $4 and $20.
• Plus: a preview of /swarm in Antigravity, which turns one prompt into a team of agents, with Carbon planning and auditing while lighter models work in parallel.

Price per token is the wrong number

Looks like a promo. It isn't. A model that deliberately burns more tokens costs more per task, even when the token price looks the same. Google's own benchmarks are framed as cost per task, not per token. That's exactly how teams should do the math. My rule from every tool comparison: credits and the follow-on price matter, not the launch discount. And the follow-on price arrives in three weeks.

So if you test Carbon in November, you're measuring at half price. If you budget off that test, you'll be paying double. That's not a knock on Google, it's a classic introductory offer. Just don't let the December bill hit your CFO's desk as a surprise.

The more interesting number is at the bottom

What grabs me isn't CursorBench. It's a figure Google mentions almost in passing: on an agentic honesty benchmark with 375 tasks, Carbon passed 92% according to Google, Argon 70%. For marketing, that's the real use case. An agent that analyzes campaign numbers or preps copy for approval is only worth as much as its willingness to say: I don't know. Whether the vendor number holds up in daily work is something you'll have to check yourself.

The catch

Every benchmark comes from Google's own presentation, some with partners. The post doesn't list countries. And "the best model" isn't a strategy. No model decides what your brand stands for or where your budget has the most leverage. AI is the ticket in, not the job.

What I'd do now

• → Take two or three real tasks from your week (a reporting analysis, a long research brief) and run Carbon against Argon. Measure cost per finished task and how often a human has to fix the output.
• → Re-run the November test numbers at December prices before anyone says "we're switching."
• → Use Carbon only where a mistake is genuinely expensive. For everything else, the smaller tool does the job.

Bottom line: The strongest model isn't the right model, and from December 1 the mistake costs twice as much.