Gemini 4 Carbon leads MathArena, but the cost gap is the other headline
Google's Gemini 4 Carbon now heads MathArena's overall recommendation table. The practical question is whether that extra benchmark performance is worth the compute bill.
MathArena's November 13 update added Carbon and Claude Fable 5.2. Its live table gives Carbon high an expected performance of 85.4% and cost of $15.70 per problem, versus GPT-6.1 Sol max at 80.6% and $1.07. Fable max is second at 82.8% and $40.20. The displayed performance intervals overlap.
These are not simple accuracy scores on an identical completed test set. MathArena combines observed scores with item-response estimates where results are missing. Expected cost averages across non-deprecated, non-Euler competitions and estimates missing costs. The figures describe these test settings, not ordinary chat pricing.
The Decoder reported the change on November 14, describing Carbon's cost as roughly fifteen times Sol's. That corroborates the reporting, not a separate reproduction of the benchmark.
Guth analysis
The useful development is the tradeoff, not another victory lap. A team solving a few exceptionally valuable problems could rationally pay more for a stronger candidate. A high-volume workflow could reach the opposite decision. Those are different purchasing questions, even when both teams look at the same leaderboard.
Our calculation from the displayed costs is approximately 14.7 times for Carbon versus Sol. That ratio should not be generalized to an API price comparison, a fixed cost to solve every problem, or a claim that the performance difference is statistically decisive. An operational evaluation should measure accepted answers, elapsed time and total spending on the user's actual workload.
Sources and citations
Revision history
Published .