ModelsCompareBest forBenchmarksStatusPricingAPI
Compare / head-to-head

Kimi K2.6vsKimi K3

Kimi K3 leads 10 of 10 shared benchmarks. Kimi K3 has the larger context window (1M tokens).

Benchmarks from cited public sources; pricing from official pages; status from official provider feeds.

Shared benchmarks
0 – 10
Kimi K3 leads
Cheaper per token
list price, input + output
Larger context
Kimi K3
1M tokens
Provider uptime (30d)
100%
Moonshot AI
Kimi K2.6
Moonshot AI · released 2026-04-20
textvision
Context
256K
Max out
Input /1M
Output /1M
Cached /1M
Scores
14 · 8 core
Kimi K3
Moonshot AI · released 2026-07-16
textvision
Context
1M
Max out
1.0M
Input /1M
$3
Output /1M
$15
Cached /1M
$0.3
Scores
59 · 8 core
Quality

Benchmark matrix

BenchmarkKimi K2.6Kimi K3ΔEdge
Reported by both · 10
GPQA Diamond90.8% epoch93.5% -2.7 ptKimi K3
AIME 202695.8% matharena96.7% matharena-0.9 ptKimi K3
FrontierMath Tier 4 v2 (Epoch AI run)25.6% epoch39% epoch-13.4 ptKimi K3
FrontierMath Tiers 1-3 v2 (Epoch AI run)57.2% epoch72.2% epoch-15 ptKimi K3
MCP-Atlas69.4% 84.2% -14.8 ptKimi K3
MCPMark-Verified72.8% 94.5% -21.7 ptKimi K3
MLS-Bench-Lite26.7% 48.3% -21.6 ptKimi K3
OTIS Mock AIME 2024-2025 (Epoch AI run)96.1% epoch97.2% epoch-1.1 ptKimi K3
ProgramBench48.3% 77.8% -29.5 ptKimi K3
SimpleQA Verified34.9% epoch50.6% epoch-15.7 ptKimi K3
Only Kimi K2.6 reports · 4
LMArena Elo1454.9 not reported
Kimi Claw 24/7 Bench42.9% not reported
Kimi Code Bench v250.9% not reported
SWE-bench Verified (Epoch AI run)76.7% epochnot reported
Only Kimi K3 reports · 47
Humanity's Last Exam (no tools)not reported43.5%
AA-Briefcasenot reported1548
AA-LCRnot reported74.7%
Agents' Last Examnot reported28.3%
APEX-Agentsnot reported41%
AutomationBenchnot reported30.8%
BabyVision (with Python)not reported85.7%
BrowseComp (1M context, no compaction)not reported90.4%
BrowseComp (context compaction)not reported91.2%
CharXiv Reasoning (no tools)not reported84.8%
CharXiv Reasoning (with Python)not reported91.3%
CorpFin v2not reported71.6%
CritPtnot reported23.4%
DeepSearchQA (F1)not reported95%
DeepSWE v1.1not reported67.3%
DeepSWE v1.1 (Kimi Code)not reported67.5%
Finance Agent v2not reported54.4%
FrontierSWEnot reported81.2%
GDPval-AA v2not reported1686
Harvey Lab-AAnot reported94.6%
Humanity's Last Exam (with tools)not reported56%
JobBenchnot reported54.3%
Kimi Code Bench 2.0not reported72.9%
Legal Research Benchnot reported44.2%
MathVision (no tools)not reported94.3%
MathVision (with Python)not reported97.8%
MMMU-Pro (no tools)not reported81.6%
MMMU-Pro (with Python)not reported83.4%
MMVUnot reported82.1%
OfficeQA Pronot reported63.3%
OmniDocBenchnot reported91.1%
OSWorld 2.0not reported58.3%
OSWorld-Verifiednot reported84.8%
PerceptionBenchnot reported58.5%
PostTrainBenchnot reported36.6%
ResearchRubricsnot reported76.2%
SaaS-Benchnot reported60.1%
SciCodenot reported58.7%
SpreadsheetBench 2not reported34.8%
SWE-Marathonnot reported42%
tau3-Bankingnot reported33.4%
Terminal-Bench 2.1not reported88.3%
Toolathlon-Verifiednot reported76.5%
Video-MME (with subtitles)not reported90%
WorldVQA ForceAnswernot reported51%
ZeroBench pass@5 (no tools)not reported23%
ZeroBench pass@5 (with Python)not reported41%
Scores tagged "epoch" or "matharena" are independent runs, used only where the lab has not published its own; ⚠ marks rows MathArena flags as released after the competition. Higher is better on every row. Δ is Kimi K2.6 minus Kimi K3 in the benchmark's own unit. "Not reported" means the lab has not published that figure; it is not a zero. ↗ opens the source.
Specs & pricing

Side by side

SpecKimi K2.6Kimi K3Edge
Context window256K tokens1M tokensKimi K3
Max output1.0M tokens
Input price / 1M$3
Output price / 1M$15
Cached input / 1M$0.3
Input + output / 1M
Lower is cheaper. List prices; batch, tool and regional fees excluded.
$18
Modalitiestext · visiontext · visionTie
Released2026-04-202026-07-16
Cited benchmark scores1459
Reliability

Moonshot AI status

All providers
More matchups

Kimi K2.6 vs …

More matchups

Kimi K3 vs …

Built by Respan
Which one wins on your data?

Public benchmarks are a starting point. Run Kimi K2.6 and Kimi K3 on your own prompts with Respan evals, or route to either through one gateway key with automatic failover.