Clarity for AI buildersSponsored by AI Routing
pricing

Gemini 4 Carbon tops reported MathArena results at roughly 15 times GPT-6.1 Sol’s cost

Gemini 4 Carbon tops reported MathArena results at roughly 15 times GPT-6.1 Sol’s cost

Gemini 4 Carbon leads MathArena in results reported by The Decoder, at roughly 15 times GPT-6.1 Sol’s cost.

According to The Decoder, Google’s Gemini 4 Carbon displaced OpenAI’s GPT-6.1 Sol at the top of the math benchmark. The supplied results put Carbon, using its “high” reasoning setting, at an expected performance of 85.4% ±3.0%. Anthropic’s Claude Fable 5.2 at “max” ranked second with 82.8% ±5.3%, followed by GPT-6.1 Sol at “max” with 80.6% ±2.7%.

The cost comparison makes the ranking less straightforward for users choosing a model. The report lists expected costs of $15.70 ±$1.33 for Carbon, $40.20 ±$2.59 for Fable and $1.07 ±$0.018 for Sol under the tested reasoning settings. Carbon’s central performance estimate is 4.8 percentage points above Sol’s, while its listed cost is about 14.7 times higher.

Fable costs about 37.6 times as much as Sol despite ranking below Carbon. The supplied text does not define the unit of work covered by these cost estimates, so they should not be treated as general API prices.

The reported performance ranges also overlap. That limits what readers can conclude from the central estimates alone: Carbon leads the supplied ranking, but the excerpt does not establish a statistically decisive advantage over either competitor. Nor does a math benchmark ranking establish superiority on unrelated tasks.

The Decoder also reports a practical limitation in Carbon’s persistence. Exceeding the time limit was its most common failure mode, and the model almost never abandoned a task. In one reported case, it launched a script that ran for more than eleven hours without checking its progress. That behavior illustrates why a higher benchmark score should be considered alongside runtime and cost: continued effort can consume resources without necessarily producing a timely answer.

The table names Moonshot AI’s Kimi K3.1 Pro at “max” as the best open model, with expected performance of 54.3% ±5.8% and expected cost of $3.70 ±$0.27. Its central performance estimate is below those of the three overall leaders, while its listed cost exceeds Sol’s. Within these reported results, the open-model designation therefore does not imply either the highest math performance or the lowest cost.

These findings are based on The Decoder’s supplied account rather than an independently inspected benchmark methodology or test log. The excerpt does not provide the task count, explain the uncertainty intervals or specify the billing basis, all of which would be needed for a fuller assessment of the performance–cost trade-off.