Skip to main content
BenchLM
Advertise

Model profile

Kimi K2.6

Moonshot AISupersededReleased Apr 20, 2026
Data verified
Superseded:Moonshot AI has released newer models in this line —Kimi K3·Kimi K2.7 Code

Kimi K2.6 by Moonshot AI scores 56.79/100 on the public leaderboard (#74 of 200), with an Arena Elo of 1461 and a 256K-token context window. API pricing is $0.95/$4 per million input/output tokens. Newer replacements: Kimi K3 and Kimi K2.7 Code.

Overall Score
56.79Public #74 of 200
Arena Elo
1461
Eligible category ranks
5of 8
Price (1M tokens)
$0.95 in / $4 out
API pricing
Speed
Not listed
Context
256K

Evidence coverage

51 of 323 tracked benchmarks are published. 32 are verified and 19 provisional. 7 of 8 categories are measured.

Updated July 23, 2026Methodology
Published / tracked
51 / 323
Verified
32
Provisional
19
Categories with evidence
7 / 8

Evidence by category

  • Agentic17 benchmarks
    Mixed evidence
  • Coding10 benchmarks
    Mixed evidence
  • Reasoning2 benchmarks
    Reported
  • Knowledge9 benchmarks
    Mixed evidence
  • Math5 benchmarks
    Verified
  • Multilingual0 benchmarks
    Not measured
  • Multimodal7 benchmarks
    Mixed evidence
  • Inst. Following1 benchmark
    Reported
Open WeightSelf-hostReasoning
Confidence:
High
base

Kimi K2.6 ranks #74 out of 200 models on the public leaderboard with an overall score of 56.79/100. It does not yet have enough sourced coverage for BenchLM's verified leaderboard. While not a frontier model, it offers specific advantages depending on the use case.

Kimi K2.6 is a open weight model with a 256K token context window. It uses explicit chain-of-thought reasoning, which typically improves performance on math and complex reasoning tasks at the cost of higher latency and token usage.

BenchLM links it directly to Kimi K2.5 as the earlier related model in that lineage. This profile currently has 51 of 323 tracked benchmarks. BenchLM only exposes non-generated benchmark rows publicly, so missing categories stay blank until a sourced evaluation is available.

Its strongest category is Mathematics (#1), while its weakest is Agentic (#98). This performance profile makes it particularly strong for mathematical reasoning, scientific computing, and quantitative analysis.

Peer position

Exact provisional scores and ranks for the closest listed peers. A score can appear before a model clears the evidence threshold for a rank, so equal scores can have different rank states.

Range 56.58–57.01

  1. Qwen3.5 397B
    Alibaba
    #7157.01
    Qwen3.5 397B is #71 with a score of 57.01.
    Compare
  2. Qwen3.5-35B-A3B
    Alibaba
    #7256.97
    Qwen3.5-35B-A3B is #72 with a score of 56.97.
    Compare
  3. GPT-5.3-Codex-Spark
    OpenAI
    #7356.91
    GPT-5.3-Codex-Spark is #73 with a score of 56.91.
    Compare
  4. Kimi K2.6Current model
    Moonshot AI
    #7456.79
    Kimi K2.6 is #74 with a score of 56.79.
  5. GPT-5.4 mini
    OpenAI
    #7556.77
    GPT-5.4 mini is #75 with a score of 56.77.
    Compare
  6. Grok 4 Fast (Reasoning)
    xAI
    #7656.59
    Grok 4 Fast (Reasoning) is #76 with a score of 56.59.
    Compare
  7. Claude Haiku 4.5
    Anthropic
    #7756.58
    Claude Haiku 4.5 is #77 with a score of 56.58.
    Compare

Category percentile

More

Relative position among models eligible for each sourced category. A higher percentile means a stronger position within that category's ranked cohort; 100 is highest.

  1. Math100%
    Eligible cohort rank #1 of 7Category score 72.3
  2. Multimodal50%
    Eligible cohort rank #15 of 29Category score 67.3
  3. Knowledge14%
    Eligible cohort rank #45 of 52Category score 56.9
  4. Coding59%
    Eligible cohort rank #51 of 122Category score 52.0
  5. Agentic18%
    Eligible cohort rank #98 of 119Category score 38.0

Category evidence

Scores and ranks appear only where this model has published benchmark evidence. Categories without displayable source records remain not measured.

Category scores, ranks, weighting, benchmark coverage, and evidence status
CategoryScore
AgenticRank #98 of 119Percentile 18thWeight 22%17 benchmarksMixed sources38.0
CodingRank #51 of 122Percentile 59thWeight 20%10 benchmarksMixed sources52.0
ReasoningWeight 17%2 benchmarksReportedScore pending
KnowledgeRank #45 of 52Percentile 14thWeight 12%9 benchmarksMixed sources56.9
MathRank #1 of 7Percentile 100thWeight 5%5 benchmarksVerified72.3
MultilingualWeight 7%0 benchmarksNot measuredNot measured
MultimodalRank #15 of 29Percentile 50thWeight 12%7 benchmarksMixed sources67.3
Inst. FollowingWeight 5%1 benchmarkReportedScore pending

Self-host vs API cost

Estimates at 50,000 req/day · 1000 tokens/req average.

Kimi K2.6
API / mo$3,713
Self-host / mo$18,221
Break-even326M/day
Model the full break-even

Chatbot Arena performance

Scroll horizontally to inspect confidence intervals and vote counts.

Chatbot Arena Elo, confidence interval, and vote count by evaluation view
ViewEloConfidence intervalVotes
Text Overall1461±4.537,711
Coding1515±7.110,419
Math1480±13.61,931
Instruction Following1455±6.512,149
Creative Writing1430±8.55,885
Multi-turn1461±8.36,537
Hard Prompts1486±5.324,325
Hard Prompts (English)1487±6.711,745
Longer Query1477±6.315,445

Benchmark Details

Rows below have a displayable published verification record. Each source link and provenance note remains in the page HTML while its category is closed. Source-unverified manual rows and generated rows stay hidden.

Agentic17 benchmarks
Terminal-Bench 2.0Provider exact
66.7%Weighted 38%
Source: MoonshotAI: Kimi K2.6 tech blogProvenance: Provider exact
OSWorld-VerifiedProvider exact
73.1%Weighted 34%
Source: MoonshotAI: Kimi K2.6 tech blogProvenance: Provider exact
BrowseCompProvider exact
83.2%Weighted 28%
Source: MoonshotAI: Kimi K2.6 tech blogProvenance: Provider exact
ToolathlonProvider exact
50%Display only
Source: MoonshotAI: Kimi K2.6 tech blogProvenance: Provider exact
MCP AtlasProvider exact
55.9%Display only
Source: MoonshotAI: Kimi K2.6 model cardProvenance: MoonshotAI reports MCPMark at 55.9 on the Kimi K2.6 launch table. BenchLM maps this to the MCP Atlas display key.
Claw-EvalBenchmark exact
62.3%Display only
Source: Claw-Eval leaderboardProvenance: Claw-Eval reports this model as kimi_k26 in the official 2026-05-09 leaderboard snapshot. BenchLM stores the primary Pass^3 value on the local Claw-Eval display key.
DeepSearchQAProvider exact
92.5%Display only
Source: MoonshotAI: Kimi K2.6 tech blogProvenance: MoonshotAI highlights DeepSearchQA as an f1-score benchmark in the Kimi K2.6 tech blog and reports 92.5 on the benchmark table.
WideResearchProvider exact
80.8%Display only
Source: MoonshotAI: Kimi K2.6 model cardProvenance: MoonshotAI reports WideSearch at 80.8 item-f1 in the agentic evaluation table. BenchLM maps this to the Wide Research display key.
AA Agentic IndexReported

Artificial Analysis Agentic Index

30.3%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
τ²-bench resultsReported

τ²-Bench Tool-Agent-User Evaluation

95.9%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
GDPval-AAReported

GDPval-AA normalized

34.5%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
GDPval-AAReported
1189Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
APEX-Agents-AAReported
28.5%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
Gert LabsBenchmark exact

Gert Labs Composite Game Benchmark

56.82%Display only
Source: Gert Labs rankingsProvenance: Gert Labs reports this composite leaderboard score in the public rankings API. BenchLM scales the source gscore from 0-1 to 0-100 and stores it as a display-only agentic benchmark.
ResearchClawBenchBenchmark exact
18.0%Display only
Source: ResearchClawBench leaderboardProvenance: ResearchClawBench reports this model as ResearchHarness (Kimi-K2.6) in the official Pass@1 leaderboard. BenchLM stores the one-decimal RADS average on the local ResearchClawBench display key and excludes it from weighted rankings.
OSWorld 2.0Benchmark exact
4.6%Display only
Source: OSWorld 2.0 paperProvenance: OSWorld 2.0 reports Kimi 2.6 single action on its 500-step main table. BenchLM stores the binary completion score and notes the corresponding partial score was 22.1%.
terminalBenchHardReported
43.9%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
Coding10 benchmarks
SWE-bench VerifiedProvider exact

Software Engineering Benchmark Verified

80.2%Weighted 16%
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
SciCodeProvider exact

Scientific Code Benchmark

52.2%Weighted 16%
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
SWE-bench ProProvider exact
58.6%Weighted 10%
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
LiveCodeBench v6Provider exact
89.6%Display only
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
SWE MultilingualProvider exact
76.7%Display only
Source: MoonshotAI: Kimi K2.6 tech blogProvenance: Provider exact
Terminal-Bench 2.0Provider exact
66.7%Display only
Source: MoonshotAI: Kimi K2.6 tech blogProvenance: Provider exact
Vibe Code BenchBenchmark exact

Vibe Code Bench v1.1

37.89%Display only
Source: Vals AI: Vibe Code Bench v1.1Provenance: Vals Vibe Code Bench v1.1 reports this exact row under kimi/kimi-k2.6-thinking; BenchLM stores it on the local vibeCodeBench key.
cursorBench31Benchmark exact
47.6%Display only
Source: Cursor evals: CursorBench 3.1Provenance: Cursor reports Kimi 2.6 at this exact CursorBench 3.1 score on its public evals page. BenchLM stores it on the Kimi K2.6 row as a display-only coding-agent benchmark.
AA Coding IndexReported

Artificial Analysis Coding Index

61.8%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-SciCodeReported

Artificial Analysis SciCode

53.5%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
Reasoning2 benchmarks
AA-LCRReported

Artificial Analysis Long Context Reasoning

69.7%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
CritPtReported

Critical Physics Tasks

8.0%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
Knowledge9 benchmarks
HLEProvider exact

Humanity's Last Exam

34.7%Weighted 45%
Source: MoonshotAI: Kimi K2.6 model cardProvenance: MoonshotAI reports HLE-Full at 34.7 in the reasoning and knowledge evaluation table.
GPQAProvider exact

Graduate-Level Google-Proof Q&A

90.5%Weighted 7%
Source: MoonshotAI: Kimi K2.6 model cardProvenance: MoonshotAI reports GPQA-Diamond at 90.5 on the Kimi K2.6 launch table. BenchLM maps that exact value to the core GPQA key as well as the display-only GPQA-Diamond row.
GPQA-DProvider exact

GPQA Diamond

90.5%Display only
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
Artificial Analysis Intelligence IndexReported
44.2%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-GPQA DiamondReported

Artificial Analysis GPQA Diamond

91.1%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-HLEReported

Artificial Analysis Humanity's Last Exam

35.9%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-Omniscience IndexReported

Artificial Analysis Omniscience Index

6.4%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-Omniscience AccuracyReported

Artificial Analysis Omniscience Accuracy

32.8%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-Omniscience Hallucination RateReported

Artificial Analysis Omniscience Hallucination Rate

39.3%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
Math5 benchmarks
FrontierMath v2 (Tiers 1-3)Benchmark exact

FrontierMath v2 Tiers 1-3

38.966%Weighted 30%
Source: Epoch AI FrontierMath v2 leaderboardProvenance: Epoch AI reports FrontierMath v2 Tiers 1-3 at 38.966% for kimi-k2.6. BenchLM selects the highest published thinking effort for the model and stores the v2 benchmark slice separately.
AIME26Provider exact

AIME 2026

96.4%Weighted 25%
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
HMMT Feb 2026Provider exact

Harvard-MIT Mathematics Tournament February 2026

92.7%Weighted 25%
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
FrontierMath v2 (Tier 4)Benchmark exact

FrontierMath v2 Tier 4

14.580%Weighted 10%
Source: Epoch AI FrontierMath v2 leaderboardProvenance: Epoch AI reports FrontierMath v2 Tier 4 at 14.58% for kimi-k2.6. BenchLM selects the highest published thinking effort for the model and stores the v2 benchmark slice separately.
MMAnswerBenchProvider exact
86.0%Display only
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
Multimodal7 benchmarks
MMMU-ProProvider exact

Massive Multi-discipline Multimodal Understanding Pro

79.4%Weighted 45%
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
CharXivProvider exact

CharXiv Reasoning

80.4%Weighted 25%
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
MMMU-Pro w/ PythonProvider exact

MMMU-Pro with Python

80.1%Display only
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
MathVisionProvider exact
87.4%Display only
Source: MoonshotAI: Kimi K2.6 model cardProvenance: Provider exact
V*Provider exact
96.9%Display only
Source: MoonshotAI: Kimi K2.6 tech blogProvenance: MoonshotAI highlights V* with python at 96.9 on the Kimi K2.6 tech blog benchmark table.
AA-MMMU-ProReported

Artificial Analysis MMMU-Pro

79.4%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
Design Arena WebsiteReported

Design Arena Website Elo

1306Display only
Source: OpenRouter model benchmarksProvenance: Display-only Design Arena Website Elo synced from OpenRouter model benchmark metadata. It is excluded from BenchLM weighted scoring.
Inst. Following1 benchmark
AA-IFBenchReported

Artificial Analysis IFBench

76.0%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.

Kimi K2.6 Family

Base entry

Related Earlier Model

Kimi K2.5

Frequently Asked Questions

How does Kimi K2.6 perform overall in AI benchmarks?

Kimi K2.6 currently ranks #74 out of 200 models on BenchLM's provisional leaderboard with an overall score of 56.79. It is created by Moonshot AI. Its published context window is 256K.

Is Kimi K2.6 good for knowledge and understanding?

Kimi K2.6 ranks #45 out of 52 models in knowledge and understanding benchmarks with an average score of 56.9. There are stronger options in this category.

Is Kimi K2.6 good for coding and programming?

Kimi K2.6 ranks #51 out of 122 models in coding and programming benchmarks with an average score of 52. There are stronger options in this category.

Is Kimi K2.6 good for mathematics?

Kimi K2.6 ranks #1 out of 7 models in mathematics benchmarks with an average score of 72.3. It is among the top performers in this category.

Is Kimi K2.6 good for reasoning and logic?

Kimi K2.6 has visible benchmark coverage in reasoning and logic, but BenchLM does not currently assign it a global category rank there.

Is Kimi K2.6 good for agentic tool use and computer tasks?

Kimi K2.6 ranks #98 out of 119 models in agentic tool use and computer tasks benchmarks with an average score of 38. There are stronger options in this category.

Is Kimi K2.6 good for multimodal and grounded tasks?

Kimi K2.6 ranks #15 out of 29 models in multimodal and grounded tasks benchmarks with an average score of 67.3. There are stronger options in this category.

Is Kimi K2.6 good for instruction following?

Kimi K2.6 has visible benchmark coverage in instruction following, but BenchLM does not currently assign it a global category rank there.

Is Kimi K2.6 open source?

Yes, Kimi K2.6 is an open weight model created by Moonshot AI, meaning it can be downloaded and run locally or fine-tuned for specific use cases.

Does Kimi K2.6 have full benchmark coverage on BenchLM?

Not yet. Kimi K2.6 currently has 51 published benchmark scores out of the 323 benchmarks BenchLM tracks. BenchLM only exposes non-generated public benchmark rows, so missing categories stay blank until a sourced evaluation is available.

What is the context window size of Kimi K2.6?

Kimi K2.6 has a published context window of 256K, which determines how much text it can process in a single interaction.

Related Resources

More from Moonshot AI LLM Pricing Comparison BenchLM Methodology The AI Race Which LLM Should I Use? Benchmark Explainers
Last updated: July 23, 2026 · Runtime metrics stay blank until BenchLM has a sourced snapshot.

Choose with this week’s evidence

Join 2,000+ readers for ranking moves, new releases, pricing changes, and the evidence behind them.

Free. One email per week.