News · High impactBack to News

Kimi K3 Benchmarks Explained: No. 1 in Frontend Arena, Still Behind Claude Fable 5 and GPT-5.6 Sol Overall

Kimi K3 leads seven of Moonshot AI's 35 published comparisons and jointly leads one more, while Arena ranks it first for frontend development and ninth in text overall. For a workload using 1M input and 200K output tokens with no cache hits, K3 costs about $6—70% less than Claude Fable 5 and roughly 45% less than GPT-5.6 Sol, but about 2.63 times the listed cost of GLM-5.2.

Kimi K3 produces three distinct benchmark signals. Moonshot AI's official table covers 35 coding, agentic, reasoning, and vision evaluations. Arena places K3 first for frontend development and ninth in text overall. Artificial Analysis gives it an Intelligence Index score of 57, ranking fourth among 189 comparable models.

The numbers support a specific assessment: K3 has entered the frontier tier, with its strongest results concentrated in frontend generation, long-horizon coding, search, and automation agents. It does not lead every broad reasoning or professional-agent evaluation, and Claude Fable 5 and GPT-5.6 Sol retain the highest scores on many tests.

Overall Conclusion From the Kimi K3 Benchmarks

Across the six-model comparison published by Moonshot AI, K3 is the sole leader on seven of 35 evaluations and jointly leads one more. Its outright wins are Program Bench, SWE Marathon, BrowseComp, DeepSearchQA, Automation Bench, SpreadsheetBench 2, and OmniDocBench. It ties Claude Fable 5 at 23.0 on ZeroBench_main.

This count describes relative placement only within Moonshot AI's table; it is not a global ranking across every available model. The capability distribution is more useful: K3 performs strongly when tasks require sustained tool use, browsing, code changes, or complete deliverables, while HLE, Toolathlon-Verified, and APEX-Agents expose meaningful gaps.

Kimi K3's Strongest Coding and Agent Benchmarks

BenchmarkKimi K3Best Listed RivalK3 Position
Program Bench77.8GPT-5.6 Sol 77.61st
Terminal-Bench 2.188.3GPT-5.6 Sol 88.82nd
FrontierSWE81.2Claude Fable 5 86.62nd
SWE Marathon42.0Claude Opus 4.8 40.01st
BrowseComp91.2GPT-5.6 Sol 90.41st
DeepSearchQA95.0Claude Fable 5 94.21st
Automation Bench30.8GPT-5.6 Sol 29.71st
SpreadsheetBench 234.8Claude Fable 5 34.71st
GDPval-AA v21668 EloClaude Fable 5 17603rd

K3's 88.3 on Terminal-Bench 2.1 is only 0.5 points behind the table leader. Its 81.2 on FrontierSWE is also above GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5, and GLM-5.2. A score of 42.0 on SWE Marathon exceeds every listed rival, supporting K3's competitiveness on longer software-engineering trajectories.

The agentic results are mixed. K3 leads BrowseComp, DeepSearchQA, and Automation Bench, but its 73.2 on Toolathlon-Verified ranks fifth among the six models. Its 37.6 on APEX-Agents also ranks fifth. Strong search and workflow execution do not guarantee consistent performance on every constrained, cross-application agent task.

Limits in Reasoning and Vision Benchmarks

The reasoning results show a clearer ceiling. K3 scores 93.5 on GPQA-Diamond, below GPT-5.6 Sol at 94.1. Its 43.5 on HLE-Full trails Claude Fable 5 at 53.3, Claude Opus 4.8 at 49.8, and GPT-5.6 Sol at 44.5. With tools, K3 rises to 56.0 but remains behind the top three listed scores.

Vision performance is broadly competitive with fewer outright wins. K3 scores 81.6 on MMMU-Pro, below GPT-5.6 Sol at 83.0. Its MathVision score of 94.3 trails both GPT-5.6 Sol and Claude Fable 5. Document vision stands out: 91.1 on OmniDocBench is the highest score in the official table.

BenchmarkKimi K3Highest Listed ScoreAssessment
GPQA-Diamond93.5GPT-5.6 Sol 94.1Close to the lead
HLE-Full43.5Claude Fable 5 53.3Material gap
HLE-Full with tools56.0Claude Fable 5 63.0Tools add 12.5 points
MMMU-Pro81.6GPT-5.6 Sol 83.0Near the front
MathVision94.3GPT-5.6 Sol 95.8Competitive
OmniDocBench91.1Kimi K3 91.11st in the table

What Third-Party Benchmarks Confirm

Arena's Code Arena WebDev Overall leaderboard dated July 16, 2026 ranks Kimi K3 first at 1679±17, ahead of Claude Fable 5 at 1631±13 and GPT-5.6 Sol at 1618±13. K3 had 1,757 votes at the time and was marked Preliminary, so the margin should be monitored as the sample grows.

On the same date, Text Arena Overall ranks K3 ninth at 1486±11, near Gemini 3 Pro and GPT-5.6 Sol xhigh. The contrast between first place in frontend development and ninth place in text indicates that K3's advantage is more pronounced in visual web generation than in open-ended text conversations.

Artificial Analysis scores K3 at 57 on its Intelligence Index, ranking fourth. It measures output speed at 62 tokens per second and time to first token at 1.99 seconds. The full Intelligence Index run generated about 130 million output tokens, close to twice the 63 million median for reasoning models in the same price tier. High intelligence therefore comes with below-median speed and high token consumption.

Five Limits on Cross-Benchmark Comparisons

Moonshot AI discloses several evaluation conditions that affect comparability:

  • Different harnesses: K3 often uses Kimi Code, Claude models commonly use Claude Code, and GPT models use Codex. The harness affects tool calls, context management, and task completion.
  • Different reasoning settings: K3 uses max reasoning effort with temperature 1.0 and top-p 1.0. GPT-5.5 uses xhigh, while other models run at their own top settings.
  • Internal evaluations are mixed in: Kimi Code Bench 2.0, DECK-Bench, and PerceptionBench include internal or in-house tests with limited external reproduction material.
  • AutomationBench has two scoring implementations: Moonshot's 30.8 and Artificial Analysis' 52.7 on AutomationBench-AA use different scoring implementations and should not be read as two measurements on one scale.
  • Open weights are still pending: As of July 17, full weights are scheduled for release by July 27. Arena and Artificial Analysis currently classify K3 as proprietary; that label can only be reassessed after the weights and license are available.

Kimi K3 Cost and Price-to-Performance

Official API pricing per million tokens is $0.30 for cache-hit input, $3 for cache-miss input, and $15 for output. The comparison below assumes one task consumes 1 million input tokens and 200,000 output tokens, with no cache hits and no tool-call or long-context surcharges.

ModelInput / MTokOutput / MTokStandard Workload Cost
Kimi K3$3.00$15.00$6.00
Claude Fable 5$10.00$50.00$20.00
GPT-5.6 Sol$5.00$30.00$11.00
GLM-5.2$1.40$4.40$2.28

Under this no-cache scenario, K3 costs 70% less than Claude Fable 5 and about 45% less than GPT-5.6 Sol. It costs about 2.63 times as much as GLM-5.2. Moonshot AI reports cache-hit rates above 90% on coding workloads. If 90% of K3's 1 million input tokens hit the cache, the same workload falls to an estimated $3.57. That cache estimate illustrates K3's own cost curve; cross-provider comparisons must also account for each provider's cache-write, cache-read, and long-context rules.

Price-to-Performance in Moonshot AI's Benchmark Table

RivalK3 Head-to-Head RecordStandard Workload CostValue Assessment
Claude Fable 512 wins, 22 losses, 1 tie$6 vs $20Fable 5 wins more evaluations overall; K3 covers selected frontier tasks at about 30% of the cost
GPT-5.6 Sol19 wins, 14 losses, 1 tie$6 vs $11K3 wins more listed comparisons and costs about 45% less
GLM-5.219 wins, 0 losses$6 vs $2.28K3 has the capability advantage; GLM-5.2 retains the lowest token cost

These records come from Moonshot AI's six-model table and remain affected by different harnesses and internal benchmarks. They indicate broad price-capability positioning but cannot replace a cost test using the same business workload and toolchain.

K3 occupies a middle price tier. It is materially cheaper than Fable 5 and Sol for near-frontier frontend development, long-horizon coding, search agents, and document-vision processing. GLM-5.2 remains cheaper when minimizing token cost matters more than preserving K3's measured capability lead. Broad reasoning, tightly bounded cross-application agents, and long-session stability still need independent testing, especially after the model weights, technical report, and larger Arena samples become available.

Next step

Keep tracking Kimi K3

Continue along the same topic.

Open entity record