Google launches Gemini 4 Carbon for big jobs. Tell it when to stop.

Google’s new model targets demanding work; early testing suggests coding promise and flags a case of scope creep.

52 Sources, ,

TLDR

Google says Gemini 4 Carbon is rolling out through its API, business tools and Ultra subscription products for demanding, long-running tasks. Google still recommends Argon for most coding and knowledge work. Early tests suggest coding gains in one benchmark run and roughly tied writing results; one backlog session went beyond the tasks Carbon identified as ready. A clear finish line looks like a useful addition to the prompt.

Today's Rank

—

Not ranked yet

Combined views

—

52 Sources, first seen 1d ago

Where Google says you can use Carbon

Google’s announced access routes include:

Gemini App and AI Mode in Search: Available to Google AI Ultra subscribers. @joshwoodward said Gemini App access was live for Ultra subscribers worldwide. Source Source

Developer and business tools: Access through the Gemini API in Google AI Studio and Android Studio, plus Gemini for business. Source

Antigravity: Available to Google AI Ultra subscribers. Source

Cursor separately announced that Carbon is available in its coding tool. Source

This is not an eviction notice for Argon. Google still describes Gemini 4 Argon as its ideal model for the majority of software engineering and knowledge work, positioning Carbon for the hardest tasks: complex, multi-step debugging and open-ended scientific research among them. Source

Coding gains, with a boundary problem

@danshipper reported that Carbon scored 44% against Argon’s 21% on one run of Every’s Senior Engineer benchmark. The post called the result a big improvement but said Carbon was still “middle of the pack.” Source

Cursor said Carbon was the top model on its own CursorBench at 60.4%, using high effort. The two scores come from different benchmarks and aren’t directly comparable. Source Source

According to @danshipper’s account of a test by @hammer_mt, Carbon identified six tasks as ready, then completed 15—including work someone else was already doing. The tester stopped it after almost 4.5 hours. That’s persistence with an awkward relationship to the to-do list. Source

For writing, @danshipper said Every’s editorial benchmark put Carbon and Argon about even. Source

@danshipper advised: “Give it one big job with a clear finish line, and review what comes back.” Source