Kimi K3 vs GPT-5.6 Luna

At a Glance

Compare
Kimi K3Moonshot AI
Intelligence, Cost, and Efficiency
IntelligenceHigher is better · MM Intelligence v2.5#17 of 4670.3 score · 3/3 sources · complete#31 of 4652.7 score · 3/3 sources · complete
CostLower is better · Published-token output estimate#30 of 44$0.194 per LiveBench case#6 of 44$0.026 per LiveBench case
EfficiencyHigher is better · MM Efficiency v1.5#21 of 3852.5 score · 3/3 sources · complete#4 of 3864.8 score · 3/3 sources · complete
Pricing and Limits
Input priceFrom · USD / 1M tokens$2.85Deepinfra · Sep 21, 2026$0.20Openai · Sep 3, 2026
Output priceFrom · USD / 1M tokens$14.25Deepinfra · Sep 21, 2026$1.20Openai · Sep 3, 2026
Context windowMaximum documented tokens1,049K1,050K
Model facts checkedAug 28, 2026View model evidence →Aug 28, 2026View model evidence →

Token prices are the lowest available sourced USD rates; input and output may use different providers. Cost ranking estimates output spend on LiveBench, not a full request bill. Ranking methodology →

Available Benchmarks

All benchmark results →
BenchmarkKimi-K3GPT-5.6 Luna
ARC-AGI-1verified-v1-ba05d69f6453 · verified_score · leader94.50100% of row best · percent · Kimi K3 (Max)90.6796% of row best · percent · GPT-5.6 Luna 2026-07-30 (Max)
ARC-AGI-2verified-v2-6c676fa3e9af · verified_score · leader60.42100% of row best · percent · Kimi K3 (Max)59.5899% of row best · percent · GPT-5.6 Luna 2026-07-30 (Max)
LMArena Agent Arenaagent-2026-09-15-d25aabda0010 · outcome_score · leader6.22100% of row best · score · Kimi K3 (Max); 95% CI [5.59393783, 6.84323085]; sessions 108615; observations 9086177; rank 8-0.4494% of row best · score · GPT 5.6 Luna (xHigh); 95% CI [-1.27310649, 0.39719069]; sessions 29186; observations 2307801; rank 27
LMArena Text Arenatext-2026-09-13-d25aabda0010 · arena_rating · leader1,472.27100% of row best · rating · kimi-k3-max; 95% CI [1467.05577758, 1477.48944277]; votes 20987; rank 231,429.8997% of row best · rating · gpt-5.6-luna-xhigh; 95% CI [1425.11549148, 1434.66063238]; votes 28547; rank 86
LiveBench2026-06-25 · overall · leader81.02100% of row best · percent · kimi-k3 · 13,647 output tokens / case77.0595% of row best · percent · gpt-5.6-luna-max · 21,799 output tokens / case
ToneBench2026-09-11-10-task-4ef099199c9c · overall_score · unverified88.19100% of row best · points · Kimi K3 · 15,023 output tokens / case84.4796% of row best · points · GPT-5.6 Luna (ultra) · 17,605 output tokens / case
Overall ResultCounted from the protocol-matched rows above5 benchmark winsOverall lead0 benchmark wins

Third-party benchmark Only like-for-like primary-publisher results are shown; raw scores, relative scores, configuration, and token spend remain visible.

Quality Versus Estimated Output Cost

Full ranking →
Efficiency FrontierLiveBench overall · output estimate
Upper-left is better
xAIZ.aiMiniMaxDeepSeekOpenAIQwenGoogle DeepMindMoonshot AIAnthropic
LiveBench quality versus score-adjusted output costEach dot is a reviewed major-model configuration and is colored by developer. Higher means a better LiveBench overall score. Farther left means lower estimated output cost after adjusting by the score. A dotted line connects the non-dominated frontier observations. When models are selected, their sourced families remain prominent, unrelated observations retain their developer colors at lower opacity, and an orange ring identifies each selected model.$0.0050$0.010$0.050$0.100$0.500$1.006873798489Grok Build 0.1GLM 5.3 FlashDeepSeek V4.1 Flash (max)GPT-5.6 Sol (max)GPT-6 Astra (max)Claude Fable 5.1Score-adjusted output cost per LiveBench case (log) →LiveBench overall →
The dotted frontier connects measured, non-dominated major-model observations. With a selection, sourced families stay prominent, unrelated observations retain their developer colors at lower opacity, and orange rings mark the selected model or models. Family lines connect models only when their sourced family and generation match. Cost is estimated from published output tokens and the lowest current USD output rate; it excludes input, caching, batch discounts, and provider-specific benchmark execution details.

Side-by-Side Facts

FieldKimi-K3GPT-5.6 Luna
DeveloperMoonshot AIOpenAI
FamilyKimi K3Gpt 5 6
ModelKimi-K3GPT-5.6 Luna
VersionKimi-K3GPT-5.6 Luna
Lifecycleactiveactive
Released2026-07-16Unknown
Knowledge cutoffUnknown2026-02-16
Input modalitiesText, ImageText, Image
Output modalitiesTextText
Context window1,049K1,050K
Total parameters2.8TUnknown
Active parameters104BUnknown
LicenseotherUnknown
Open weightsYesNo
API availableYesYes
Self-hostableYesNo
Provider accessDeepinfra (Standard), Fireworks Ai (Standard), Hugging Face (Standard), Openrouter (Standard), Together Ai (Standard)Openai (Standard), Openrouter (Standard)
Capabilitieschat, generation, reasoningchat, generation, reasoning, tools

Kimi K3 Capabilities

chatgenerationreasoning
Serving providers5
Canonical IDmoonshotai/Kimi-K3

GPT-5.6 Luna Capabilities

chatgenerationreasoningtools
Serving providers2
Canonical IDopenai/gpt-5.6-luna

Primary Evidence

Sources and Freshness

Questions

Kimi K3 vs GPT-5.6 Luna FAQs

Is Kimi K3 or GPT-5.6 Luna better for coding?+

This comparison does not currently contain a protocol-matched coding benchmark for both Kimi K3 and GPT-5.6 Luna, so Model Markets cannot name a coding leader from pricing, context size, or capability labels alone.

Which is cheaper, Kimi K3 or GPT-5.6 Luna?+

Kimi K3 is $2.85 and GPT-5.6 Luna is $0.20 per million tokens, so GPT-5.6 Luna is cheaper on this metric. Kimi K3 is $14.25 and GPT-5.6 Luna is $1.20 per million tokens, so GPT-5.6 Luna is cheaper on this metric.

Which has a larger context window, Kimi K3 or GPT-5.6 Luna?+

GPT-5.6 Luna has the larger sourced context window. Kimi K3 supports 1,049K and GPT-5.6 Luna supports 1,050K.

Which performs better in benchmarks, Kimi K3 or GPT-5.6 Luna?+

Kimi K3 leads the current overall benchmark count. The result uses 5 protocol-matched benchmarks from 3 publishers; it is not a universal quality score.

Can Kimi K3 or GPT-5.6 Luna be self-hosted?+

Kimi K3 is the only model in this pair currently marked as self-hostable. Kimi K3 is open weight; GPT-5.6 Luna is not marked open weight.

Can Kimi K3 and GPT-5.6 Luna understand images?+

Kimi K3 is documented with image input; GPT-5.6 Luna is documented with image input. This reflects supported input modalities, not vision quality.

Which can generate longer answers, Kimi K3 or GPT-5.6 Luna?+

Neither has a larger sourced maximum output. Kimi K3 is — and GPT-5.6 Luna is 128K.

Do Kimi K3 and GPT-5.6 Luna support reasoning and tool use?+

Kimi K3: reasoning and image input. GPT-5.6 Luna: reasoning, tool calling, and image input. Feature support does not establish relative quality.

Which is available from more inference providers, Kimi K3 or GPT-5.6 Luna?+

Kimi K3 has 5 sourced provider routes; GPT-5.6 Luna has 2, so Kimi K3 has broader tracked availability.

Which offers better value, Kimi K3 or GPT-5.6 Luna?+

There is no universal value winner. Compare the input and output prices above with the matched benchmark result for your workload: cheaper tokens can be offset by different quality, token usage, latency, or provider availability.

Send Feedback