Google's Gemini 4 Carbon Trails Anthropic's Opus on Accuracy
Google's Gemini 4 Carbon cut its false-claim rate from 36% to 18% in Theo Browne's review, but most rivals, including Anthropic's Opus, fabricate at 0–1%.
Gemini 4 Carbon's false-claim rate fell from 36% for the previous Gemini model to 18%, according to developer Theo Browne's review.
Browne said most competing models, including Anthropic's Opus, fabricate at 0–1%.
In Artificial Analysis data, Carbon's average task took 22 minutes on low reasoning effort, while Opus took 3 minutes 1 second on low.
The introductory price of $2 per million input tokens and $10 per million output tokens doubles on December 1, 2026.
Browne argues that cost per task, not token price, is the figure that matters.