Picture a math exam where one student hands in the best paper in the room, and the bill for their scrap paper is fifteen times bigger than the bill for the student who finished third. That is roughly what the MathArena leaderboard looks like this week.
Google’s Gemini 4 Carbon now sits at number one. According to The Decoder, it is the first non-OpenAI model to lead the board for a while. Here is what the scores say, what the price tag really measures, and why a 15x cost gap does not make Carbon a bad deal or a good one on its own.
The scoreboard, as of 15 November 2026
MathArena is run by SRI Lab at ETH Zurich and INSAIT, with a grant from Epoch AI. It tests models on fresh problems from recent contests such as AIME 2026, USAMO 2026 and the 2025 Putnam, so the answers are less likely to be sitting in a training set. Each model runs four times on each problem.
Here are the top of the table and a few useful neighbors:
| Rank | Model (reasoning setting) | Maker | Expected performance | Expected cost |
|---|---|---|---|---|
| 1 | Gemini 4 Carbon (high) | 85.4% ±3.0% | $15.70 | |
| 2 | Claude Fable 5.2 (max) | Anthropic | 82.8% ±5.3% | $40.20 |
| 3 | GPT-6.1 Sol (max) | OpenAI | 80.6% ±2.7% | $1.07 |
| 6 | Claude Opus 5.5 (high) | Anthropic | 75.3% ±3.2% | $3.92 |
| 12 | Gemini 4 Argon (high) | 62.1% ±9.9% | $11.49 | |
| Best open model | Kimi K3.1 Pro (max) | Moonshot AI | 54.3% ±5.8% | $3.70 |
The “15 times” in The Decoder’s headline is simple division: $15.70 divided by $1.07 is about 14.7. Claude Fable 5.2 works out to roughly 38 times Sol’s cost, which The Decoder rounds to roughly 40.
What “expected cost” means
Here is the part that trips people up. MathArena says its leaderboards show the average cost of a single run on one problem, in US dollars. So $15.70 covers one attempt at one problem, on average, while GPT-6.1 Sol’s equivalent figure is $1.07.
Why so much? Google’s own launch post says Carbon uses more tokens than its sibling Argon to explore problems and run extended verification. Tokens are the chunks of text a model reads and writes, and you pay for each one. A model that thinks longer writes a longer bill.
Google’s published API price is $2 per million input tokens and $10 per million output tokens as an introductory rate, rising to $4 and $20 after 30 November 2026. That matches Argon’s introductory price, according to Google. Yet MathArena lists Carbon (high) at $15.70 against $11.49 for Argon (high). With the same per-token price, the gap points to how many tokens each model spends.
The model that would not quit
The Decoder, citing a post on X by MathArena’s Jasper Dekoninck, reports that Carbon almost never gives up. In one case it ran a script for more than eleven hours without checking on its progress. Its most common failure was running past the time limit.
That is either admirable persistence or the behavior of a colleague who disappears into a spreadsheet for a full workday. Google, for its part, says internal teams have seen Carbon run for days on multi-step engineering tasks like finding and patching bugs in large codebases.
Before you crown anyone
Three caveats, all visible in the table itself.
- The margins overlap. By our arithmetic from the published figures, Carbon’s range runs from about 82.4% to 88.4% and Sol’s from about 77.9% to 83.3%. Fable 5.2’s range is wider still and overlaps both. Carbon leads, but the evidence is thin.
- The settings differ. Carbon is listed at “high” reasoning, while Fable 5.2 and Sol are listed at “max”. The cost and score comparisons are not like for like.
- Math is one job. On other benchmarks the cost story flips. Google says Carbon ranked first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at more than 2x lower cost per task than the next-best external model. Those are Google’s own claims, so treat them as vendor-reported. SemiAnalysis’s GDPx, a 206-task test across coding, finance, legal work and biology, put Carbon second behind Claude Opus 5.5, at just over $16 per task against about $51 for Opus 5.5 and about $54 for GPT-6 Astra, according to 9to5Google.
What this means for you
If you are not solving olympiad problems for a living, the 85.4% matters less than the shape of the trade-off. Carbon looks built for long, hard jobs where you want a model to keep checking its own work. Sol looks like the cheap option for the same math. Which one is the better value depends on whether your task needs the extra thinking.
Carbon is already reachable. Google says it is available through the Gemini API for paying developers, and rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity. If you want to call Gemini models from code, our guide to calling Gemini 4 Hydrogen through the Gemini API covers the setup. For the Anthropic side of this leaderboard, see our explainer on Claude Fable 5.2.
A practical habit: when you compare models, ask for cost per correct answer on your own tasks, not cost per token or a leaderboard rank.
What happens next
Expect the ranking to move. MathArena keeps adding new competitions, and its entries carry release dates, with Carbon listed on 10 November and Fable 5.2 on 9 November. Google’s introductory pricing also ends on 30 November, which will change the cost column for anyone running Carbon through the API. The Decoder also notes that OpenAI has models, not yet publicly available, that reportedly solved harder problems, so the crown may not stay put for long.
Sources
Links last checked on 15 November 2026.
- The Decoder: Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol
- MathArena: homepage and leaderboard
- MathArena: models table
- Google: Gemini 4 Carbon is now available
- Google: Gemini 4 Carbon, our next frontier model
- 9to5Google: Gemini 4 Carbon coverage
- Android Headlines: Gemini 4 Carbon benchmark and cost





Write a response