Moonshot AI Open weights: moonshotai/Kimi-K3Reasoning model

Kimi K3

Reasoning effort

Kimi K3 is a Moonshot AI open-weights reasoning model, released 16 Jul 2026. Its best configuration (max reasoning effort) ranks #39 of 756 on the BenchLeader Index at 65.6 ±3.1, in the upper half. The ± is the point: 101 other configurations score within that range, so they and this one cannot be told apart on quality alone — price and speed are what separate them. It scores highest in composite (79) and lowest in knowledge (63). At $6.00 per million tokens blended it is among the most expensive ranked models. Output speed of 34 tokens per second puts it in the slowest quarter, with a first answer in 63.8 s. It has been measured at 5 reasoning-effort settings; this summary describes the best-scoring one, and the tabs above switch between them. Last measured 2 Oct 2026.

Blended price: $6.00/M; $3.00 in · $15.00 out
Output speed: 34 tok/s; measured by Artificial Analysis
First answer: 64 s; first token 1.27 s
Context: 1.0M
Full answer: 79 s; median, reasoning included
Cost per run: $2.00; one full Intelligence Index run
Released: 16 Jul 2026

How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.

Overall index66
Reasoning67
Coding65
Agents & tools66
Maths64
Knowledge63
Human preference69
Multimodal64
Long context70
Composite79

Reasoning-effort configurations

The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Not stated” collects results from publishers that did not say which setting they used; for a reasoning model that is usually its thinking mode, but we do not assume it. Pick a setting here or at the top of the page to see its price, speed and category scores.

Effort	Index	Rank	Speed	First answer	Chat reply cost	Categories
low	58.3	#149	36 tok/s	61 s	$0.0057	Agents & tools 71 · Coding 56 · Composite 64 · Knowledge 64 · Long context 65 · Maths 54 · Multimodal 62 · Reasoning 49
high	–	–	43 tok/s	1.27 s	$0.0057	Maths 65 · Reasoning 62
maxbest	65.6	#39	34 tok/s	64 s	$0.0057	Agents & tools 66 · Coding 65 · Composite 79 · Human preference 69 · Knowledge 63 · Long context 70 · Maths 64 · Multimodal 64 · Reasoning 67
thinking	–	–	43 tok/s	1.27 s	$0.0057	Maths 65
not stated	61.8	#88	43 tok/s	1.27 s	$0.0057	Agents & tools 59 · Coding 63 · Composite 62 · Human preference 71 · Knowledge 65 · Maths 73

How its index has moved

8 Sept 2026 to 5 Oct 2026

max
not stated
low

The index is recomputed from scratch every day, so a line moves when a new benchmark result lands, when a publisher revises a score, or when the models it is normalised against change. Early movement usually means the score is still settling.

Benchmark results

One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.

Metadata

Benchmark	low	high	max	thinking	not stated	Source	Trend
AA Openness Indexnot in index	–	–	38.9#30	–	–	Artificial Analysis

Reasoning

Benchmark	low	high	max	thinking	not stated	Source
GPQA Diamond	84.8%#87	91.9%#26	93.1%#20	–	–	Epoch AI Benchmarking Hub
SimpleBench	–	–	60.7%#28	–	–	SimpleBench
LMArena Hard Prompts	–	–	1517#9	–	–	LMArena
LiveBench Reasoningnot in index	–	–	–	–	90.7%#9	LiveBench
GPQA Diamond (AA)not in index	84.2%#142	–	93.5%#14	–	–	Artificial Analysis
Humanity's Last Exam (AA)not in index	25.0%#167	–	46.9%#39	–	–	Artificial Analysis
GPQA Diamond (Vals)not in index	–	–	–	–	92.9%#12	Vals AI
ARC-AGI-1	65.7%#127	86.7%#84	94.5%#35	–	–	ARC Prize
ARC-AGI-2	12.4%#132	55.0%#86	60.4%#74	–	–	ARC Prize
CritPt	3.1%#157	–	23.4%#46	–	–	Artificial Analysis
Chess Puzzlesnot in index	20.0%#85	25.0%#65	39.0%#31	–	–	Epoch AI Benchmarking Hub
Mystery Game Puzzlesnot in index	12.0%#97	–	26.0%#38	–	–	Epoch AI Benchmarking Hub
LMCAnot in index	–	–	52.7%#40	–	–	Epoch AI Benchmarking Hub
DTBenchnot in index	–	–	91.2%#58	–	–	Epoch AI Benchmarking Hub
ForecastBenchnot in index	–	–	61.1%#15	–	–	ForecastBench

Coding

Benchmark	low	high	max	thinking	not stated	Source
SciCode	51.2%#89	–	59.5%#12	–	58.7%#20	SciCode
WeirdML	–	–	82.6%#18	–	–	WeirdML
FrontierCode	–	–	–	–	44.2%#14	Cognition
LMArena Coding	–	–	1541#9	–	–	LMArena
LMArena WebDev	–	–	1658#13	–	–	LMArena
LiveBench Codingnot in index	–	–	–	–	81.5%#14	LiveBench
SciCode (AA)not in index	52.7%#75	–	59.5%#12	–	–	Artificial Analysis
LiveCodeBench	–	–	87.2%#17	–	–	Vals AI
IOI	–	–	48.9%#23	–	–	Vals AI
SWE-bench (Vals)not in index	–	–	–	–	93.4%#8	Vals AI
Code Migrationnot in index	–	–	16.1%#46	–	–	Vals AI
ProgramBenchnot in index	–	–	–	–	2.0%#6	Vals AI
Vibe Code Bench 1-100not in index	–	–	18.2%#10	–	–	Vals AI
Vibe Code Bench v1.1not in index	–	–	–	–	85.0%#13	Vals AI
Surface Evolver Benchnot in index	–	–	93.0%#4	–	95.0%#1	Epoch AI Benchmarking Hub
DeepSWEnot in index	–	–	68.5%#19	–	–	Epoch AI Benchmarking Hub
ALE-Benchnot in index	–	–	1524.5#17	–	–	Epoch AI Benchmarking Hub
FrontierSWEnot in index	–	–	25.9%#11	–	–	Epoch AI Benchmarking Hub
Terminal-Bench 4.0 (AA)not in index	12.6%#67	–	12.6%#67	–	–	Artificial Analysis
Terminal-Bench 2.1 (AA)not in index	82.4%#38	–	85.0%#26	–	–	Artificial Analysis

Agents & tools

Benchmark	low	high	max	thinking	not stated	Source
APEX-Agents	–	–	–	–	50.6%#27	Mercor
LMArena Agent	–	–	4.2#14	–	–	LMArena
LiveBench Agentic Codingnot in index	–	–	–	–	62.2%#9	LiveBench
Terminal-Bench 2.1 (Vals)	–	–	–	–	80.9%#9	Vals AI
MCP Atlas	–	–	82.3%#8	–	–	Scale AI SEAL
GDPval (AA)not in index	30.9%#117	–	52.0%#37	–	–	Artificial Analysis
τ²-Bench Banking (AA)	41.6%#25	–	46.0%#11	–	–	Artificial Analysis
ITBench SRE (AA)not in index	–	–	47.7%#9	–	–	Artificial Analysis
Analyst Agent (AA)not in index	–	–	38.8%#18	–	–	Artificial Analysis
APEX-Agents (AA)not in index	–	–	41.3%#3	–	–	Artificial Analysis
CyberBenchnot in index	–	–	75.2%#6	–	–	Vals AI
Finance Agent v2not in index	–	–	–	–	54.4%#22	Vals AI
Harvey's Legal Agent Benchmarknot in index	–	–	10.8%#12	–	–	Vals AI
Legal Research Benchnot in index	–	–	–	–	44.2%#16	Vals AI
Tax Agent Benchnot in index	–	–	68.7%#11	–	–	Vals AI
Terminal-Bench 4.0 (Vals)not in index	–	–	19.7%#19	–	–	Vals AI
Terminal-Bench Sciencenot in index	–	–	2.9%#23	–	–	Vals AI
Time Horizon Index: KSPnot in index	–	–	10.5%#7	–	–	Vals AI
PostTrainBenchnot in index	–	–	–	–	32.0%#6	Epoch AI Benchmarking Hub
GBAEvalnot in index	–	–	–	–	48.4%#9	Epoch AI Benchmarking Hub
Vending-Bench 2not in index	–	–	–	–	5165.0#27	Andon Labs
AutomationBenchnot in index	–	–	58.3%#24	–	–	Artificial Analysis
GDP.pdfnot in index	–	–	22.0%#22	–	–	Artificial Analysis
Harvey LABnot in index	–	–	94.6%#2	–	–	Harvey
EnterpriseOps-Gymnot in index	–	–	45.3%#9	–	–	Artificial Analysis
AA-Briefcasenot in index	–	–	1501#25	–	–	Artificial Analysis

Maths

Benchmark	low	high	max	thinking	not stated	Source
FrontierMath Tiers 1–3	–	–	72.2%#22	–	–	Epoch AI Benchmarking Hub
FrontierMath Tier 4	–	–	39.0%#29	–	–	Epoch AI Benchmarking Hub
OTIS Mock AIME	68.9%#133	93.3%#51	97.2%#30	–	–	Epoch AI Benchmarking Hub
ProofBench	–	–	–	–	87.0%#9	Vals AI
LiveBench Mathematicsnot in index	–	–	–	–	84.4%#51	LiveBench
AIME 2026	–	–	–	96.7%#6	–	MathArena
HMMT February 2026	–	–	–	97.0%#3	–	MathArena
MathArena Apex	–	–	–	65.6%#4	–	MathArena
LMArena Maths	–	–	1501#16	–	–	LMArena

Knowledge

Benchmark	low	high	max	thinking	not stated	Source
SimpleQA Verified	–	–	50.6%#24	–	–	Epoch AI Benchmarking Hub
LiveBench Data Analysisnot in index	–	–	–	–	78.7%#20	LiveBench
AA-Omniscience	3.9#104	–	19.7#66	–	–	Artificial Analysis
MMLU-Pro	–	–	88.0%#22	–	–	Vals AI
LegalBench	–	–	–	–	86.0%#15	Vals AI
CorpFin	–	–	–	–	71.6%#3	Vals AI
TaxEval	–	–	–	–	75.7%#12	Vals AI
BioMysteryBenchnot in index	–	–	71.5%#9	–	–	Vals AI
Excel Modeling Benchmarknot in index	–	–	66.4%#18	–	–	Vals AI
MedCodenot in index	–	–	–	–	48.9%#25	Vals AI
MedScribenot in index	–	–	–	–	88.0%#13	Vals AI
Public Benefits Benchnot in index	–	–	68.3%#9	–	–	Vals AI
GDP.pdfnot in index	–	–	19.0%#28	–	–	Epoch AI Benchmarking Hub
AA-Omniscience: accuracynot in index	45.8%#84	–	47.6%#73	–	–	Artificial Analysis
AA-Omniscience: non-hallucinationnot in index	22.9%#244	–	46.8%#123	–	–	Artificial Analysis

Instruction following

Benchmark	low	high	max	thinking	not stated	Source
LiveBench Languagenot in index	–	–	–	–	85.5%#11	LiveBench
LiveBench Instruction Followingnot in index	–	–	–	–	71.4%#23	LiveBench
LMArena Instruction Followingnot in index	–	–	1488#14	–	–	LMArena

Human preference

Benchmark	low	high	max	thinking	not stated	Source
LMArena Text	–	–	1488#16	–	–	LMArena
EQ-Bench 4	–	–	–	–	1339#3	EQ-Bench
LMArena Creative Writingnot in index	–	–	1460#26	–	–	LMArena
LMArena Multi-turnnot in index	–	–	1500#9	–	–	LMArena
LMArena Longer Queriesnot in index	–	–	1504#10	–	–	LMArena

Multimodal

Benchmark	low	high	max	thinking	not stated	Source
MMMU-Pro	78.5%#69	–	80.5%#43	–	–	Artificial Analysis
MMMU-Pro (Vals)not in index	–	–	–	–	88.2%#11	Vals AI
MortgageTaxnot in index	–	–	–	–	66.3%#34	Vals AI
SAGEnot in index	–	–	–	–	54.3%#4	Vals AI
Blueprint-Bench 2not in index	–	–	–	–	29.5%#17	Epoch AI Benchmarking Hub

Long context

Benchmark	low	high	max	thinking	not stated	Source	Trend
AA-LCR	79.3%#89	–	88.7%#1	–	–	Artificial Analysis
MLCRnot in index	–	–	38.3%#12	–	–	Artificial Analysis

Composite

Benchmark	low	high	max	thinking	not stated	Source
Epoch Capabilities Indexnot in index	–	–	–	–	157.6#13	Epoch AI Benchmarking Hub
LiveBench	–	–	–	–	79.2%#14	LiveBench
AA Intelligence Index	30.1#109	–	43.6#42	–	–	Artificial Analysis
Vals Indexnot in index	–	–	–	–	50.7#26	Vals AI
Vals Multimodal Indexnot in index	–	–	–	–	73.4%#3	Vals AI

Safety & honesty

Benchmark	low	high	max	thinking	not stated	Source	Trend
FORTRESSnot in index	–	–	26.6%#29	–	27.3%#28	Scale AI SEAL

Where it wins

Benchmarks where this configuration ranks in the top five of every configuration measured.

#1AA-LCR88.7% · of 521

Where to run it

Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.

Provider	Speed	First token	Input $/M	Output $/M	Context	Quantisation
inference.net Fast	83 tok/s	0.79 s	$3.50	$15.00	250k	fp4
AkashML	81 tok/s	1.21 s	$3.00	$15.00	1.0M	fp4
Parasail	74 tok/s	1.12 s	$3.00	$15.00	1.0M	fp4
Fireworks Fast	74 tok/s	1.08 s	$4.50	$22.50	1.0M	–
Relace	72 tok/s	1.02 s	$0.660	$13.00	1.0M	fp4
Makora	72 tok/s	0.93 s	$2.04	$12.75	1.0M	–
Baseten	64 tok/s	2.21 s	$3.00	$15.00	1.0M	fp8
Amazon Bedrock	58 tok/s	2.21 s	$3.00	$15.00	1.0M	–
Together	53 tok/s	0.98 s	$2.70	$13.50	1.0M	–
Fireworks	50 tok/s	2.55 s	$4.50	$22.50	1.0M	–
Phala	49 tok/s	2.13 s	$1.95	$9.75	1.0M	–
Modal	48 tok/s	3.14 s	$3.00	$15.00	1.0M	mxfp4
Wafer	44 tok/s	0.75 s	$2.80	$14.00	1.0M	–
inference.net	43 tok/s	1.22 s	$0.670	$14.00	1.0M	fp4
Wafer	41 tok/s	0.67 s	$1.15	$15.00	1.0M	–
DeepInfra	40 tok/s	0.90 s	$2.85	$14.25	1.0M	mxfp4
Alibaba Cloud Int.	37 tok/s	3.46 s	$3.45	$17.25	1.0M	–
DigitalOcean	33 tok/s	1.59 s	$2.55	$12.95	1.0M	–
Morph	31 tok/s	1.26 s	$2.42	$11.46	1.0M	fp8
Moonshot AI	27 tok/s	5.21 s	$3.00	$15.00	1.0M	mxfp4
Fireworks	25 tok/s	2.35 s	$3.00	$15.00	1.0M	–
Chutes	20 tok/s	4.10 s	$3.00	$15.00	1.0M	mxfp4
Decart	1 tok/s	160 s	$1.80	$9.00	1.0M	mxfp4

Price history

Listed price per 1M tokens over time, as recorded by OpenRouter for the provider with the longest history.

input outputnow $1.00 in · $10.50 out

What a task costs

Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache at $0.300 per 1M. Reasoning tokens are not modelled.

Workload	Tokens in / out	Cost	With caching	Time
Chat reply	400 / 300	$0.0057	$0.0049	1.2 min
Summarise a 30-page report	12,000 / 600	$0.045	$0.021	1.4 min
Code edit	6,000 / 1,500	$0.041	$0.028	1.8 min
Agentic coding session	60,000 / 4,000	$0.240	$0.118	3.0 min
Structured extraction	2,000 / 200	$0.0090	$0.0050	1.2 min