A first that is real, and narrower than it sounds

OpenAI models have held the top of MathArena for a long stretch, so a change at number one is news on its own. The benchmark's operator put it plainly: "for the first time in a very long time, we have a non-OpenAI model as #1 on MathArena."

The benchmark is a reasonable place for that to happen. MathArena does not recycle a fixed test set that models can absorb during training. It uses mathematics competitions and research problems released after the models were built, and it runs each problem four times per model. A lead there is harder to dismiss than a lead on a static board.

The table argues with the headline

Read the whole row, though, and first place gets less certain. Carbon's 85.4% carries a reported range of ±3.0 points. Fable 5.2's 82.8% carries ±5.3. The 2.6-point gap between them is smaller than either range, and the intervals overlap, so the results published so far do not settle which of the two is actually better at this kind of mathematics. Sol at 80.6% ±2.7 sits close enough behind that its interval overlaps Carbon's range as well.

There is a second qualification, and it is the more unusual one. Carbon has not yet been run across the full benchmark. Dekoninck said cost limited the team to the ArXivLean evaluation, with LeanLean to follow; Carbon's own page lists measured results on ArXivMath and BrokenArXiv sets. Where a model has no measured result, MathArena's methodology fills the gap with an item-response theory estimate drawn from other models and problems, and that estimate counts in the 85.4%. Some of the number at the top of the board is therefore measured work, and some of it is the benchmark's best estimate of work not yet done. MathArena does not state what share of Carbon's figure the estimates make up, and the same policy applies to any model with missing results.

Fifteen times the price of third place

MathArena publishes expected cost next to score, which is where this result becomes a purchasing question. The table's expected cost for one run at one problem is $15.70 with Carbon and $40.20 with Fable 5.2. Sol does the same run for $1.07. Carbon's lead over Sol is 4.8 points of expected performance bought at roughly fifteen times the cost per problem, and Fable's second place costs roughly thirty-eight times Sol's rate (both ratios are ALFW calculations from the table's figures).

The Decoder reports that Carbon's most common failure mode on the benchmark is exceeding the time limit. Dekoninck made the cost point himself in a follow-up post: under these reasoning settings, the extra performance "does come at significantly higher costs." These results use the reasoning settings listed for each model.

For a research team solving hard problems where one correct answer is worth a great deal, that trade can be rational. For anyone routing ordinary hard-problem workloads, the table's practical reading is different: the cheapest model in the top three is within five points of the top, and the two models above it cost roughly fifteen and thirty-eight times more per attempt.

What would settle it

Two updates would firm up the ranking. The first is Carbon run on the remaining evaluations, replacing estimates with measurements; the operator has already said LeanLean is to follow. The second is a direct uncertainty analysis of the differences between the models, rather than a comparison of separate intervals. Until then, the accurate summary is the one the benchmark's own operator gave it: a non-OpenAI model leads MathArena, at a cost he flagged in the same thread.

Sources

  1. [1] MathArena leaderboard and methodology (primary source for scores, intervals, costs and the estimation policy)Read source
  2. [2] Jasper Dekoninck's announcement and follow-ups on X, 13–14 November 2026, as quoted by AI OriensRead source
  3. [3] The Decoder, "Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol", 14 November 2026 (failure-mode detail)Read source