Mercury 3 wins more areas, narrowly.
Scores updated · 7 tests both models report · How we compare
Where each one wins
Tests won in each of the two areas where both have results. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
Mercury 3 costs 88% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where Mercury 3 pulls ahead
- Abstract visual puzzles that people can solveARC-AGI-2+50points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+6.6points ahead
- Graduate-level biology, physics and chemistry questionsGPQA Diamond+1.3points ahead
Where Kimi K2.5 pulls ahead
- Keeps track of context across a multi-turn chatMulti-Challenge+1points ahead
Every test, side by side
All 7 tests both models report. The winning score is in its model's colour; marks a score checked independently.
ReasoningMercury 3
- ARC-AGI-2Mercury 3 by 5011.861.8+50
- GPQA DiamondMercury 3 by 1.387.989.2+1.3
Following instructionsEven
- IFBenchMercury 3 by 6.670.276.8+6.6
- Multi-ChallengeKimi K2.5 by 161.460.4+1
Other results3 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- ARC-AGI-1Mercury 3 by 23.265.388.5+23.2
- τ³-Bench BankingMercury 3 by 1614.230.2+16
- DeepSearchQA (F1)Kimi K2.5 by 5.877.171.3+5.8
Questions people ask
Which is better, Kimi K2.5 or Mercury 3?
Mercury 3 wins one of the two areas where both have results: reasoning. Kimi K2.5 wins none. They are level on following instructions.
Which is cheaper?
Kimi K2.5 costs $0.60 per million input tokens and $3.00 per million output tokens; Mercury 3 costs $0.10 and $0.35. That makes Mercury 3 about 88% cheaper for the same work.
How do you compare the two?
We use the 7 benchmark tests both models have published scores on. The verdict counts the 4 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 3 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.