Moonshot AI Open weights: moonshotai/Kimi-K2.5Reasoning model

Kimi K2.5

Name: Kimi K2.5 benchmark results
Creator: BenchLeader

Best configuration ranks #162 of 610 on the BenchLeader Index at 56.1 ±3.0. Last measured 13 Feb 2026stale: no new result in six months. Released 27 Jan 2026.

Blended price: $1.20/M; $0.600 in · $3.00 out
Output speed: 37 tok/s
First answer: 1.64 s
Context: 262k

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

Overall index56
Reasoning53
Coding53
Agents & tools48
Maths65
Knowledge52
Instruction following62
Multimodal55
Long context65
Composite59

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. “Default” means the publisher did not say which setting was used.

Effort	Index	Rank	Speed	First answer	Chat reply cost	Categories
no reasoning	52.7	#241	37 tok/s	1.64 s	$0.0011	Agents & tools 50 · Composite 54 · Instruction following 46 · Knowledge 58 · Long context 60 · Multimodal 58
high	–	–	37 tok/s	1.64 s	$0.0010	Coding 62
thinking	55.3	#178	37 tok/s	1.64 s	$0.0010	Agents & tools 32 · Coding 55 · Human preference 64 · Knowledge 60 · Maths 53 · Multimodal 62 · Reasoning 68
defaultbest	56.1	#162	37 tok/s	1.64 s	$0.0011	Agents & tools 48 · Coding 53 · Composite 59 · Instruction following 62 · Knowledge 52 · Long context 65 · Maths 65 · Multimodal 55 · Reasoning 53

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.

Reasoning

Benchmark	no reasoning	high	thinking	default	Source
GPQA Diamond	–	–	–	87.6%#60	Epoch AI Benchmarking Hub
Humanity's Last Exam	–	–	–	24.4%#14	Scale AI / CAIS
SimpleBench	–	–	–	46.8%#48	SimpleBench
LMArena Hard Prompts	–	–	1472#68	–	LMArena
GPQA Diamond (AA)not in index	78.9%#200	–	–	87.9%#92	Artificial Analysis
Humanity's Last Exam (AA)not in index	13.2%#210	–	–	30.7%#99	Artificial Analysis
GPQA Diamond (Vals)not in index	–	–	84.1%#52	–	Vals AI
Kagi LLM Benchmark	–	–	78.5%#10	63.8%#40	Kagi LLM Benchmark
ARC-AGI-1	–	–	–	65.3%#103	ARC Prize
ARC-AGI-2	–	–	–	11.8%#104	ARC Prize

Coding

Benchmark	no reasoning	high	thinking	default	Source
SWE-bench Verified (Epoch)	–	–	–	73.8%#17	Epoch AI Benchmarking Hub
SciCode	–	–	–	49.0%#66	SciCode
WeirdML	–	–	–	45.6%#81	WeirdML
LMArena Coding	–	–	1502#60	–	LMArena
LMArena WebDev	–	–	1436#59	–	LMArena
LiveCodeBench	–	–	83.9%#44	–	Vals AI
IOI	–	–	17.7%#26	–	Vals AI
SWE-bench (Vals)not in index	–	–	70.0%#58	–	Vals AI
SWE-bench Verified (bash only)	–	70.8%#11	–	–	SWE-bench
SWE-bench Verified (any scaffold)not in index	–	70.8%#15	–	–	SWE-bench

Agents & tools

Benchmark	no reasoning	high	thinking	default	Source
Terminal-Bench	–	–	–	43.2%#34	Terminal-Bench
APEX-Agents	–	–	–	14.4%#47	Mercor
Terminal-Bench Hard	18.9%#161	–	–	34.9%#80	Artificial Analysis
τ²-Bench Telecom (AA)not in index	81.3%#114	–	–	95.9%#16	Artificial Analysis
Terminal-Bench 2.1 (Vals)	–	–	42.0%#54	–	Vals AI
MCP Atlas	–	–	–	64.4%#22	Scale AI SEAL

Maths

Benchmark	no reasoning	high	thinking	default	Source
OTIS Mock AIME	–	–	–	92.2%#52	Epoch AI Benchmarking Hub
AIME (Vals)	–	–	95.6%#8	–	Vals AI
AIME 2026	–	–	95.8%#13	–	MathArena
HMMT February 2026	–	–	87.1%#19	–	MathArena
MathArena Apex	–	–	8.8%#23	–	MathArena

Knowledge

Benchmark	no reasoning	high	thinking	default	Source
SimpleQA Verified	–	–	–	34.3%#43	Epoch AI Benchmarking Hub
AA-Omniscience	-13.8#158	–	–	-7.3#124	Artificial Analysis
MMLU-Pro	–	–	85.9%#50	–	Vals AI
CorpFin	–	–	68.3%#9	–	Vals AI
TaxEval	–	–	74.2%#39	–	Vals AI
MedQA	–	–	94.4%#17	–	Vals AI
PRBench Finance	–	–	–	46.5%#14	Scale AI SEAL
PRBench Legal	–	–	–	43.8%#18	Scale AI SEAL
MultiNRC	–	–	–	35.2%#21	Scale AI SEAL

Instruction following

Benchmark	no reasoning	high	thinking	default	Source
IFBench	43.7%#216	–	–	70.2%#69	Artificial Analysis
MultiChallenge	–	–	–	61.4%#9	Scale AI SEAL
TutorBench	–	–	–	54.6%#6	Scale AI SEAL

Human preference

Benchmark	no reasoning	high	thinking	default	Source	Trend
LMArena Text	–	–	1451#65	–	LMArena

Multimodal

Benchmark	no reasoning	high	thinking	default	Source
LMArena Vision	–	–	1267#36	–	LMArena
MMMU-Pro	73.1%#99	–	–	75.4%#74	Artificial Analysis
VISTA	–	–	–	41.9%#37	Scale AI SEAL

Long context

Benchmark	no reasoning	high	thinking	default	Source	Trend
Fiction.LiveBench 120k	–	–	–	78.1%#6	Fiction.live
AA-LCR	67.3%#182	–	–	78.0%#82	Artificial Analysis

Composite

Benchmark	no reasoning	high	thinking	default	Source
Epoch Capabilities Indexnot in index	–	–	–	148.0#51	Epoch AI Benchmarking Hub
AA Intelligence Index	19.4#180	–	–	23.5#131	Artificial Analysis
Vals Indexnot in index	–	–	26.3#44	–	Vals AI

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
Venice	104 tok/s	1.33 s	$0.532	$3.33	256k	–
Amazon Bedrock	44 tok/s	1.74 s	$0.600	$3.00	262k	–
AtlasCloud	37 tok/s	0.95 s	$0.490	$2.50	262k	int4
Phala	37 tok/s	1.83 s	$0.600	$3.00	262k	–
NovitaAI	36 tok/s	1.54 s	$0.570	$2.85	262k	–
SiliconFlow	19 tok/s	3.27 s	$0.450	$2.25	262k	int4

Price history

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

input outputnow $0.450 in · $2.25 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 $0.100 per 1M. Reasoning tokens are not modelled.

Workload	Tokens in / out	Cost	With caching	Time
Chat reply	400 / 300	$0.0011	$0.0010	9.7 s
Summarise a 30-page report	12,000 / 600	$0.0090	$0.0045	17.9 s
Code edit	6,000 / 1,500	$0.0081	$0.0059	42.2 s
Agentic coding session	60,000 / 4,000	$0.048	$0.025	1.8 min
Structured extraction	2,000 / 200	$0.0018	$0.0011	7.0 s