I noted Google’s announcement of Gemini 4 Carbon as a model for long, multi-step agent work. The performance and cost figures are Google’s own claims, and the post leaves open whether the next AI race will favor intelligence or low-cost persistence.
Gemini 4 Carbon launched on 10 November.
Google says Gemini 4 Carbon is built for long, multi-step work.
- The post names complex coding, debugging, scientific research, and cybersecurity as intended work, and says Google calls it its strongest model for long-horizon agent work.
The post reports that Gemini 4 Carbon ranks first on CursorBench 4.0.
The post reports first place on Perplexity's WANDR and Zapier's AutomationBench, at more than twice the cost efficiency per task of the next best model.
The post reports first place on Perplexity's WANDR and Zapier's AutomationBench.
The post claims more than twice the cost efficiency per task of the next best model on these independent benchmarks.
On Google's new MKA benchmark, the post reports that Gemini 4 Carbon ranks second at about a sixth of the top model's cost.
The post reports a 0% prompt injection success score for Carbon on Gray Swan's benchmark.
Google reportedly used Carbon to formally verify parts of Signal's cryptographic protocol, work the post says used to take expert teams years.
The post gives Gemini 4 Carbon's introductory price as $2 per million input tokens, matching Argon.
I treat the performance claims cautiously: the post says these are Google's own numbers.
I left open whether the next race is about the smartest model, or the one that can keep going cheapest.