DataLearner logo

Kimi K2.6vsKimi K2.5

Across 20 shared benchmarks, Kimi K2.6 leads overall: Kimi K2.6 wins 17, Kimi K2.5 wins 3, with 0 ties and an average score difference of +13.91.

Moonshot AI
Kimi K2.6

Moonshot AI · 2026-04-20 · Reasoning model

Moonshot AI
Kimi K2.5

Moonshot AI · 2026-01-27 · Multimodal model

Kimi K2.617 wins(85%)(15%)3 winsKimi K2.5

Benchmark scores

Grouped by capability, sorted by largest gap within each. 20 shared benchmarks.

Agent Level Benchmark

Kimi K2.6 3/3
BenchmarkKimi K2.6Kimi K2.5Diff
Terminal Bench Hard37.9059 / 244Normal (With Tools)18.90150 / 244Normal (With Tools)+19
τ²-Bench - Telecom93.9043 / 264Normal (With Tools)81.30108 / 264Normal (With Tools)+12.60
τ³-Banking23.3082 / 164Thinking (With Tools)14.20113 / 164Thinking (With Tools)+9.10

Coding and Software Engineer

Kimi K2.6 3/3
BenchmarkKimi K2.6Kimi K2.5Diff
SWE-Bench Pro - Public58.6018 / 62Thinking (With Tools)50.7047 / 62Thinking (With Tools)+7.90
LiveCodeBench89.6011 / 250Thinking (No Tools)8529 / 250Thinking (No Tools)+4.60
SWE-bench Verified80.2014 / 116Thinking (With Tools)76.8030 / 116Thinking (With Tools)+3.40

General Knowledge

Kimi K2.6 3/3
BenchmarkKimi K2.6Kimi K2.5Diff
HLE19.60290 / 563Normal (No Tools) · Text only13.20357 / 563Normal (No Tools) · Text only+6.40
LiveBench70.5434 / 117Thinking (No Tools)69.0741 / 117Thinking (No Tools)+1.47
CritPt1.40135 / 200Normal (No Tools)0.60168 / 200Normal (No Tools)+0.80

AI Agent - Tool Usage

Kimi K2.6 2/2
BenchmarkKimi K2.6Kimi K2.5Diff
Terminal-Bench 2.165.9094 / 192Thinking (With Tools)45.70137 / 192Thinking (With Tools)+20.20
Terminal Bench 2.066.7010 / 48Thinking (With Tools)50.8035 / 48Thinking (With Tools)+15.90

Math and Reasoning

Kimi K2.6 2/2
BenchmarkKimi K2.6Kimi K2.5Diff
IMO-AnswerBench8610 / 24Thinking (No Tools)81.8018 / 24Thinking (No Tools)+4.20
AIME 202696.409 / 29Thinking (No Tools)92.5021 / 29Thinking (No Tools)+3.90

AI Agent - Information Search

Kimi K2.6 1/1
BenchmarkKimi K2.6Kimi K2.5Diff
BrowseComp83.2016 / 57Thinking (With Tools + Internet)60.6038 / 57Thinking (With Tools + Internet)+22.60

Claw-style Agent Evaluation

Kimi K2.5 1/1
BenchmarkKimi K2.6Kimi K2.5Diff
Claw Bench80.9019 / 29Thinking (With Tools)81.7018 / 29Thinking (With Tools)-0.80

General Evaluation

Kimi K2.5 1/1
BenchmarkKimi K2.6Kimi K2.5Diff
GPQA Diamond78.80249 / 462Normal (No Tools)78.90248 / 462Normal (No Tools)-0.10

Instruction Following

Kimi K2.6 1/1
BenchmarkKimi K2.6Kimi K2.5Diff
IF Bench44.30184 / 282Normal (No Tools)43.70188 / 282Normal (No Tools)+0.60

Long Context

Kimi K2.6 1/1
BenchmarkKimi K2.6Kimi K2.5Diff
AA-LCR69.70106 / 170Normal (No Tools)67.30113 / 170Normal (No Tools)+2.40

Text Embedding

Kimi K2.5 1/1
BenchmarkKimi K2.6Kimi K2.5Diff
Context Arena51.8886 / 126Normal (No Tools)53.3383 / 126Normal (No Tools)-1.45

Writing and Creative Capabilities

Kimi K2.6 1/1
BenchmarkKimi K2.6Kimi K2.5Diff
Creative Writing1,72125 / 106Normal (No Tools)1,57643 / 106Normal (No Tools)+145.50

Specs

FieldKimi K2.6Kimi K2.5
PublisherMoonshot AIMoonshot AI
Release date2026-04-202026-01-27
Model typeReasoning modelMultimodal model
ArchitectureMoEMoE
Parameters1T1T
Context length256K256K
Max outputNot available16K

API pricing

Prices use DataLearner records when available; missing fields are not inferred.

ItemKimi K2.6Kimi K2.5
Text input$0.95 / 1M tokens$0.6 / 1M tokens
Text output$4 / 1M tokens$3 / 1M tokens
Cache read$0.16 / 1M tokens$0.1 / 1M tokens
Cache write$0.95 / 1M tokensNot public

Summary

  • Kimi K2.6leads in:Agent Level Benchmark (3/3), Coding and Software Engineer (3/3), General Knowledge (3/3), AI Agent - Tool Usage (2/2), Math and Reasoning (2/2), AI Agent - Information Search (1/1), Instruction Following (1/1), Long Context (1/1), Writing and Creative Capabilities (1/1)
  • Kimi K2.5leads in:Claw-style Agent Evaluation (1/1), General Evaluation (1/1), Text Embedding (1/1)

On average across the 20 shared benchmarks, Kimi K2.6 scores 13.91 higher.

Largest single-benchmark gap: Creative Writing — Kimi K2.6 1,721 vs Kimi K2.5 1,576 (+145.50).

Page generated from structured model, pricing and benchmark records. No real-time LLM is used to write the prose.