TopicGoogle Gemini 4
Gemini 4 Carbon Tops MathArena at 85.4%, but Costs 15x GPT-6.1 Sol
MathArena updated its leaderboard on Nov 13: Google's Gemini 4 Carbon tops the math benchmark at 85.4%, Anthropic's Claude Fable 5.2 is second at 82.8%, and OpenAI's GPT-6.1 Sol drops to third at 80.6%.

What happened
On Nov 13, the independent math benchmark MathArena added results for Gemini 4 Carbon and Claude Fable 5.2, and the top spot changed hands: OpenAI's GPT-6.1 Sol, the previous leader, dropped to third. The leaderboard ranks models by expected performance, combining measured results with statistical estimates.
According to the leaderboard cited by The Decoder on Nov 14, Gemini 4 Carbon (high) scores 85.4% expected performance (±3.0 points) at $15.70 per problem (±$1.33); Claude Fable 5.2 (max) scores 82.8% (±5.3) at $40.20; GPT-6.1 Sol (max) scores 80.6% (±2.7) at $1.07. Carbon was released on Nov 10, Fable 5.2 on Nov 9, and Sol on Sep 29.
The Decoder notes that Carbon's most common failure mode is exceeding the time limit, and that it almost never gives up on a task: in one run, it launched a script that ran for more than eleven hours without checking on its progress. The best open model is Moonshot AI's Kimi K3.1 Pro at 54.3%, or $3.70 per problem.
Key facts
- First: Gemini 4 Carbon (high), 85.4% expected performance (±3.0), $15.70 per problem (±$1.33), released Nov 10
- Second: Claude Fable 5.2 (max), 82.8% (±5.3), $40.20 per problem, released Nov 9
- Third: GPT-6.1 Sol (max), 80.6% (±2.7), $1.07 per problem, released Sep 29
- Cost: At the reasoning settings used in testing, Carbon costs about 15 times as much per problem as Sol, and Fable 5.2 about 40 times as much
- Failure mode: Carbon's most common failure is exceeding the time limit, and it almost never gives up; in one run a script ran for more than eleven hours
- Error ranges: The three leaders' error ranges overlap: Carbon 82.4% to 88.4%, Fable 5.2 77.5% to 88.1%, Sol 77.9% to 83.3%
- Best open model: Kimi K3.1 Pro, 54.3% (±5.8), $3.70 per problem
Why it matters
The leader costs $15.70 per problem, about 15 times the third-place model, and the leaders' error ranges overlap, so first place does not mean the best value.
Follow HIMIAO Horizon
No sign-up · Unsubscribe any timeNotifications
A ping on your phone or computer when the brief is out. No email needed.
Comments
Related
- Nov 3SemiAnalysis Launches GDPx: Even the Best Model Scores 12%
- Oct 29Gemini 4 Carbon vs GPT-6.1 Terra: Two New Models in One Week, Double the Price Gap and an 8.4-Point Index Gap
- Sep 30Google Unveils Gemini 4 'Argon': Its First New-Generation Flagship Since Gemini 3
- Oct 21OpenAI's Frontier Models Come to Google Cloud
- Oct 17Elorian Launches First Visual Reasoning Model Glimpse: Reasoning With Images
Today
Nov 17 Tuesday- Brief
- GitHubGitHub Weekly Vol.6: 7 Projects Worth a Look This Week
Tested or read — weeklyAll →
- Quiz
- Call
Latest news
All →- Nov 17OpenAI Launches Science Blog as GPT-6 Astra Leads Proof of Candidate Lattice Standard Model Definition
- Nov 17Apple Smart Glasses Said to Debut at WWDC 2027 With No Display, Leaning on Siri and iPhone
- Nov 17AI Labs Race to Hire Counterintelligence Staff: Anthropic 33 This Year, OpenAI 22
- Nov 17Trump Backs Huang's Return to Nvidia: 'He Learned a Big Lesson'
- Nov 16Kuo: OpenAI and Anthropic to Buy About 26M Google TPUs in 2028–29
- Nov 16Huawei Cloud Launches Ascend-Native Embodied World Model CloudRobo-WM1.2
™