Kimi K2.5 vs Kimi K3
Wins 0 of 7 areas
—
Wins 7 of 7 areas
Coding · Agents · Reasoning · Facts · Images and charts · Math · Long documents
Kimi K3 is the stronger all-rounder.Kimi K2.5 is cheaper.
Scores updated · 38 tests both models report · How we compare
Where each one wins
Tests won in each of the seven 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.
- AgentsCarrying out multi-step tasks on its own06Kimi K36 of 6 tests
- ReasoningHard problems that need careful thinking04Kimi K34 of 4 tests
- CodingWriting and fixing software03Kimi K33 of 3 tests
- FactsGetting facts right instead of making them up03Kimi K33 of 3 tests
- Images and chartsUnderstanding pictures, charts and video02Kimi K32 of 2 tests
- MathCompetition and research-level math01Kimi K31 of 2 tests · 1 tie
- Long documentsFinding answers in very long texts01Kimi K31 of 1 test
Long documents rests on a single test.
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.
Kimi K2.5 costs 80% 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 Kimi K3 pulls ahead
- Abstract visual puzzles that people can solveARC-AGI-2+48.6points ahead
- Command-line tasks in a real terminalTerminal-Bench 2.1+39.3points ahead
- Real work tasks from 44 professionsGDPVal+34.5points ahead
Where Kimi K2.5 pulls ahead
No clear win on a test scored out of 100.
Every test, side by side
All 38 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingKimi K3
- Terminal-Bench 2.1Kimi K3 by 39.345.785+39.3
- LMArena · WebDevKimi K3 by 219 rating points14361655+219 rating
- SciCodeKimi K3 by 10.54959.5+10.5
AgentsKimi K3
- GDPValKimi K3 by 34.517.251.7+34.5
- τ-Bench V3 · BankingKimi K3 by 31.814.246+31.8
- AA ApexAgentsKimi K3 by 29.811.541.3+29.8
- OSWorld-VerifiedKimi K3 by 21.563.384.8+21.5
- MCP AtlasKimi K3 by 20.463.884.2+20.4
- BrowseCompKimi K3 by 16.374.991.2+16.3
ReasoningKimi K3
- ARC-AGI-2Kimi K3 by 48.611.860.4+48.6
- CritPtKimi K3 by 20.33.123.4+20.3
- Humanity's Last ExamKimi K3 by 16.230.746.9+16.2
- GPQA DiamondKimi K3 by 5.687.993.5+5.6
FactsKimi K3
- SimpleQA VerifiedKimi K3 by 16.334.350.6+16.3
- AA-Omniscience · Non-hallucinationKimi K3 by 12.534.346.8+12.5
- AA-Omniscience · AccuracyKimi K3 by 12.435.247.6+12.4
Images and chartsKimi K3
- CharXiv (RQ)Kimi K3 by 13.877.591.3+13.8
- MMMU-ProKimi K3 by 5.175.480.5+5.1
Other results17 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.
- GDPval-AA v2Kimi K3 by 673 rating points10091682+673 rating
- BabyVisionKimi K3 by 49.236.585.7+49.2
- ToolathlonKimi K3 by 45.427.873.2+45.4
- CyberGymKimi K3 by 38.741.380+38.7
- ZeroBenchKimi K3 by 301141+30
- Artificial Analysis Coding IndexKimi K3 by 29.446.876.2+29.4
- ARC-AGI-1Kimi K3 by 29.265.394.5+29.2
- AA Agentic IndexKimi K3 by 28.921.750.6+28.9
- AA-OmniscienceKimi K3 by 27-7.319.7+27
- NL2RepoKimi K3 by 263258+26
- AA IntelligenceKimi K3 by 20.123.543.6+20.1
- DeepSearchQA (F1)Kimi K3 by 17.977.195+17.9
- ResearchRubricsKimi K3 by 16.759.576.2+16.7
- MATH-VisionKimi K3 by 13.684.297.8+13.6
- matharena_visual_math_overallKimi K3 by 10.780.691.3+10.7
- HLE (with tools)Kimi K3 by 9.650.259.8+9.6
- WorldVQAKimi K3 by 4.746.351+4.7
Questions people ask
Which is better, Kimi K2.5 or Kimi K3?
Kimi K3 wins all seven areas where both have results: coding, agents, reasoning, facts, images and charts, math and long documents. Kimi K2.5 wins none, but costs 80% less.
Which is better for coding?
Kimi K3. It wins 3 of the 3 coding tests both models report; Kimi K2.5 wins none.
Which is cheaper?
Kimi K2.5 costs $0.60 per million input tokens and $3.00 per million output tokens; Kimi K3 costs $3.00 and $15.00. That makes Kimi K2.5 about 80% cheaper for the same work.
How do you compare the two?
We use the 38 benchmark tests both models have published scores on. The verdict counts the 21 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 17 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.