Kimi K2.6 Thinking

Kimi K2.6 Thinking is the reasoning-optimized K2.6 variant for deeper multi-step planning and execution. It is tuned for long-horizon coding and design workflows, including complex orchestration across many specialized sub-agents and autonomous end-to-end output generation.

Reasoning
Vision
Tool Calling
Parallel Tool Calls

Added Apr 16, 2026

Model weights

Pricing

Auto routing · per 1M tokens

Input: $0.44
Output: $1.86
Cache read: $0.074

Compare provider prices

Specifications

Context window: 262.1K
Max output: 65.5K
Parameters: 1T / 32B; Total / active
Avg output (7d): 4.1K tokens; Longer than 99% of models

Benchmarks

Artificial Analysis

LMArena

Vectara

Sourced from Artificial Analysis.

Intelligence Index

27.0

Better than 81% of models compared

Coding Index

61.8

Better than 74% of models compared

Agentic Index

20.5

Better than 58% of models compared

Agentic work

AutomationBench-AA

Workflow automation with guardrail penalties

13.0%

Better than 36% of models compared

Harvey LAB-AA

Legal agentic work criterion pass rate

84.1%

Better than 33% of models compared

AA-Briefcase

Agentic knowledge work (Elo)

819 Elo

Better than 35% of models compared

GDPval-AA v2

Economically valuable tasks (Elo)

1040 Elo

Better than 49% of models compared

Document reasoning

GDP.pdf

Professional PDF reasoning: all-pass rate

13.0%

Better than 52% of models compared

AA-LCR v1.1

Long context reasoning with updated grading

81.0%

Better than 90% of models compared

Reasoning

HLE

Humanity's Last Exam

37.5%

Better than 85% of models compared

IFBench

Instruction-following benchmark

76.0%

Better than 95% of models compared

Coding

Terminal-Bench v4.0

Practical coding and terminal tasks

0.5%

Better than 35% of models compared

SciCode

Python programming for scientific computing

51.5%

Better than 59% of models compared

Legacy benchmarks

GPQA Diamond (legacy)

Graduate-level scientific reasoning

91.1%

Better than 91% of models compared

Terminal-Bench Hard (legacy)

Agentic coding and terminal use

43.9%

Better than 91% of models compared

T²-Bench Telecom (legacy)

Conversational AI agents in dual-control scenarios

95.9%

Better than 96% of models compared

AA-LCR (unversioned / legacy)

Long context reasoning evaluation

81.0%

Better than 89% of models compared

Last updated Oct 14, 2026