Google’s Gemini 4 Carbon has knocked OpenAI’s GPT-6.1 Sol from first place on the MathArena maths benchmark, with Anthropic’s Claude Fable 5.2 second.
Carbon’s cost per benchmark run is about 15 times GPT-6.1 Sol’s at the reasoning settings used in testing, and Fable 5.2’s is about 38 times.
The Decoder reports that Carbon is the first non-OpenAI model in a while to top MathArena. MathArena added Fable 5.2 and Carbon on 13 November. The table lists the four entries The Decoder reports, with the reasoning setting each model ran at.
| Provider | Model (reasoning setting) | Expected performance | Expected cost per run |
|---|---|---|---|
| Gemini 4 Carbon (high) | 85.4% ±3.0% | $15.70 ±$1.33 | |
| Anthropic | Claude Fable 5.2 (max) | 82.8% ±5.3% | $40.20 ±$2.59 |
| OpenAI | GPT-6.1 Sol (max) | 80.6% ±2.7% | $1.07 ±$0.018 |
| Moonshot AI | Kimi K3.1 Pro (max) | 54.3% ±5.8% | $3.70 ±$0.27 |
Kimi K3.1 Pro is the best open model on the list, some 31 points behind Carbon.
The lead sits inside the error margins
Carbon’s 85.4% ±3.0% overlaps with both rivals’ ranges, so the ranking is closer than the table order suggests. Fable 5.2’s ±5.3% is the widest of the three.
Why the cost gap is so big
The 15 times figure is the cost of a whole benchmark run, not a price per token. Yahoo reports that Carbon’s API price at launch was $2 per million input tokens and $10 per million output tokens. That matches OpenAI’s list price for GPT-6 Sol. GPT-6.1 Sol’s token price isn’t in any of the sources.
The Decoder points to how much work Carbon does per task. Google pitches Carbon mainly at cybersecurity and long-running tasks. The Decoder says Carbon’s most common failure is exceeding the time limit, and it almost never gives up. In one run it ran a script for more than eleven hours without checking progress.
Availability and pricing
Google made Carbon available through the API on 10 November, as Tbreak reported when it covered Gemini 4 Hydrogen in Google Search. Yahoo reports the standard rate is $4 input and $20 output per million tokens once the introductory period ends.
On Google’s own MKA benchmark, a visual office-task test, ppc.land reports Fable 5.2 first and Carbon second. Carbon’s cost per task is about 16% of Fable’s ($5.63 against $35.28, per Google’s technical report). Tbreak’s May article on Google’s AI plans in the UAE gives the local subscription prices.
