Hy3-preview vs Kimi K2.6
Wins 1 of 7 areas
Math
Wins 5 of 7 areas
Coding · Reasoning · Facts · Long documents · Following instructions
Kimi K2.6 is the stronger all-rounder.Hy3-preview is cheaper and better at math.
Scores updated · 30 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.
- CodingWriting and fixing software04Kimi K2.64 of 7 tests · 3 ties
- ReasoningHard problems that need careful thinking03Kimi K2.63 of 3 tests
- FactsGetting facts right instead of making them up01Kimi K2.61 of 2 tests · 1 tie
- Long documentsFinding answers in very long texts01Kimi K2.61 of 1 test
- Following instructionsDoing exactly what it is asked01Kimi K2.61 of 1 test
- MathCompetition and research-level math10Hy3-preview1 of 1 test
- AgentsCarrying out multi-step tasks on its own22Even2 each · 2 ties
Math, long documents and following instructions rest on a single test each.
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.
Hy3-preview costs 86% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Kimi K2.6 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+33.6points ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+28points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+12.1points ahead
Where Hy3-preview pulls ahead
- Multi-step tasks using many real software toolsMCP Atlas+11points ahead
- Olympiad math problems with short answersIMOAnswerBench+4points ahead
- Real work tasks from 44 professionsGDPVal+1.3points ahead
Every test, side by side
All 30 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingKimi K2.6
- Terminal-Bench HardKimi K2.6 by 12.131.843.9+12.1
- SciCodeKimi K2.6 by 2.948.651.5+2.9
- SWE-bench VerifiedKimi K2.6 by 2.27880.2+2.2
- Terminal-Bench 2.1Kimi K2.6 by 1.564.465.9+1.5
- SWE-bench Multilingualtie75.876.7tie
- SWE-bench Protie57.958.6tie
- LMArena · WebDevtie15081509tie
AgentsEven
- MCP AtlasHy3-preview by 1179.168.1+11
- AA ApexAgentsKimi K2.6 by 2.925.628.5+2.9
- BrowseCompKimi K2.6 by 2.184.286.3+2.1
- GDPValHy3-preview by 1.328.327+1.3
- τ-Bench V3 · Bankingtie22.923.3tie
- Terminal-Bench 4.0tie0.50.5tie
ReasoningKimi K2.6
- Humanity's Last ExamKimi K2.6 by 433.537.5+4
- CritPtKimi K2.6 by 3.14.98+3.1
- GPQA DiamondKimi K2.6 by 1.489.791.1+1.4
FactsKimi K2.6
- AA-Omniscience · Non-hallucinationKimi K2.6 by 33.625.959.5+33.6
- AA-Omniscience · Accuracytie3232.6tie
Other results9 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.
- τ²-Bench Telecom (AA run)Kimi K2.6 by 28.467.595.9+28.4
- AA-OmniscienceKimi K2.6 by 23.8-18.55.3+23.8
- Claw Eval (pass@3)Kimi K2.6 by 12.468.580.9+12.4
- DeepSearchQA (F1)Hy3-preview by 89183+8
- WideSearchKimi K2.6 by 4.476.480.8+4.4
- AA Agentic IndexHy3-preview by 3.525.622.1+3.5
- Artificial Analysis Coding IndexKimi K2.6 by 358.861.8+3
- AA IntelligenceKimi K2.6 by 1.725.327+1.7
- ToolathlonKimi K2.6 by 1.548.550+1.5
Questions people ask
Which is better, Hy3-preview or Kimi K2.6?
Kimi K2.6 wins five of the seven areas where both have results: coding, reasoning, facts, long documents and following instructions. Hy3-preview wins math, and costs 86% less. They are level on agents.
Which is better for coding?
Kimi K2.6. It wins 4 of the 7 coding tests both models report; Hy3-preview wins none, and 3 are ties.
Which is cheaper?
Hy3-preview costs $0.14 per million input tokens and $0.56 per million output tokens; Kimi K2.6 costs $0.95 and $4.00. That makes Hy3-preview about 86% cheaper for the same work.
How do you compare the two?
We use the 30 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 9 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.