HighlightsEdition

ArtificialAnalysisStateofAI

Q22025

HighlightsReport

FllpilblAIeTidsbscribcribers

ArtificialAnalysisisaleading,andindependentAIbenchmarkingandinsightsprovider.WesupportengineersandcompaniestounderstandAIcapabilitiesandmakecriticaldecisionsabouttheirAIstrategy.

Ourdata,insightsandpublicationsaregroundedinourcomprehensivebenchmarkingofAItechnologiesandusecases.ThisincludeseverythingfromhourlyperformancetestingoflanguagemodelAPIstomillionsofvotesinourcrowd-sourcedarenas.

Ourpublicwebsite,artificialanalysis.ai,iswidelyreferencedbycompaniesleadinginnovationinAI.Todiscussthisreport,ourpublications,orourservices,pleasegetintouchatcontact@artificialanalysis.ai.

ArtificialAnalysisAITrendsSubscription:ComprehensiveAImarketintelligenceforenterprisedecision-makingfromtheleadingAIbenchmarkingcompany

A|QuarterlyStateofAIReport

|EnterpriseAgentsReport

C|AIAdoptionSurvey

|Databooks&API

ThedefinitivequarterlyupdateonAI

2025istheyearofagents–our

Real-worldadoptioninsightsfrom

Directaccesstotheindustry'smost

marketdevelopments

overviewofwhatmattersmost

thosebuildinganddeployingAI

comprehensivedata

?Emergingtrendsateachlayerofthe

?Comprehensiveanalysisofkey

?Enterpriseusecasepatterns

?ComprehensiveAIperformance

AIstack:hardware,infrastructure,models

?Marketmapsandperformance

rankingsforhundredsofkeyplayers

?StateofAI:China-detailed

agentcategories:coding,deep

research,computeruse,customersupport,sales

?What'sworkingnow:whereagentsaredrivingrealproductivity

?Enterpriseadoptionbenchmarks

?Developerprioritiesandpainpoints

?Model,inferenceandhardwareproviderdemandbyindustry

data-sourcedataforallouranalysis

?Intelligence,performance,cost,surveydataandmore

?ExceldatabooksandAPIaccess

benchmarkingoftopAIlabsin

?Implicationsforreal-world

China

deployment

E|QuarterlyAITrendsWorkshop

Connectmarketintelligencetoyourstrategicpriorities

?LivebriefingwithArtificialAnalysisresearchteam(90minutes,optional)

?What'sworkingnow:insightsfromleadingSFstartupsandenterprises

?Deepdivestailoredtoyourbusinesspriorities(e.g.,codingagentbestpractices,inferenceeconomics,upcomingchips)

F|OngoingTeamAccess

Directaccesstoourresearchteamforsupportandclarifications

?Supportforuseofreportsanddata,includingqueriesonsourcesandmethodology

?Clarificationandexplanationofanalyses

?Limitedtomax90minutesperquarter,ratesavailableforfurthersupport

ArtificialAnalysisis

trustedbytheleadingAIindustryplayersand

publications

OverviewofArtificialAiITrendsSubscription

ThisistheHighlightsVersionoftheQuarterlyStateofAIReportforQ22025,thePremiumVersionisavailabletosubscribersofourAITrendsSubscription

HighlightsVersion(This)PremiumVersion(AITrendsSubscription)

IndustryoverviewandmarketmapofkeyplayersandstrategiesacrosstheAIvaluechain

OverviewoffrontiermodelsrankedbytheArtificialAnalysisIntelligenceIndexandoverviewofemergingtrends

Synthesisofemergingtrendsforimage,videoandspeechmodelsandmarketmaps

SynthesisofemergingtrendsforacceleratorsincludingcasestudycomparingNVIDIAB200andNVIDIAH200using

ArtificialAnalysisSystemLoadTest

IncludeseverythingintheHighlightsVersionplus:

Detailedinsightsacrossnewlanguagemodelreleases(incl.analysisofleadingopenweightsoptions)

Detailedanalysisandcasestudiesoutliningemergingtrendsforlanguagemodelsacrosspricing,performanceandfeatures

Analysisoffrontierimagegenerationmodelsandtrends

(incl.texttoimageandimageediting)

Analysisoffrontiervideogenerationmodelsandtrends(incl.texttovideoandimagetovideo)

Analysisoffrontierspeechmodelsandtrends(incl.texttospeechandspeechtotext)

Emergingmarkettrendsforaccelerators,includingdetailedanalysiscomparingNVIDIAH200andNVIDIAB200

AttachedSeparateReport:EnterpriseAgentsReport

coveringcomprehensiveanalysisofkeyagentcategoriesandimplicationsforreal-worlddeployment

Feelfreetogetintouchwithusat

subscriptions@artificialanalysis.ai

tolearnmoreabouttheArtificialAnalysisAITrendsSubscription

ArtificialAnalysisStateofAIQ22025

ThestoryofAIinQ22025revealsanindustryhittingitsstrideafteryearsoffoundationaldevelopment.WearewitnessinganewphasewhereinnovationsacrosstheAIstackarematuringandconvergingtowardsimpactinghoweveryorganizationoperates.

Today'smodelsdemonstratesignificantintelligencegainswhile

becomingmorecost-effectiveandfasterthanever.Agenticworkflowsaremovingfrompromisingexperimentstoproductionreality,with

codingagentsproliferatingacrossdevelopmentteams.Meanwhile,thecompetitivelandscapecontinuestoevolve,withChineseAIlabsdemonstratingremarkableleadershipinbothlanguageandvideo

capabilities.

ProducedbyArtificialAnalysis,anindependentbenchmarkingandinsightsfirmtrustedacrosstheAIvaluechain,thisQ22025reportisdesignedtoinforminvestment,product,andpolicydecisionsinanincreasinglyAI-nativeworld.

Formoredetails,contactusat

founders@artificialanalysis.ai

5

-MicahHill-SmithandGeorgeCameron,FoundersofArtificialAnalysis

Contents

1.IndustryOverview

OverviewofmarketmovementsandtrendsbykeyplayersintheAIindustry

2.LanguageModels

Trendsinfrontierlanguagemodels,including

hybridmodels,costandefficiencyimprovements

3.ImageandVideo

Trendsinfrontierimageandvideomodels

includinganoverviewoftheleadingmodelsinArtificialAnalysisImageandVideoArenas

4.SpeechandAudio

TrendsacrossnewspeechmodelsandanoverviewofnewandleadingmodelsintheArtificialAnalysisSpeechArena

5.Accelerators

OverviewoftheAIacceleratormarketincludingmarkettrends,available

acceleratorsandverticalintegrationbyselectchipmakers

01

7

Toolsandconnections

enablesmartworkflow

integration

Nativeconnectionsandtoolsinchatinterfacesarenowshiftingworkloadstoagenticapproaches

Languagemodelscontinuetobecomemoreintelligent

MajorAIlabshaveallcontinued

tomakesubstantialgainsin

intelligence,costefficiencyand

speed

Codingagentsrapidly

proliferateacross

developmentworkflows

Q2saw12majorcodingagent

launches,includingfrommajorlabs

5majortrendshave

shapedtheStateofAI

acrossQ22025

Videomodelssee

breakthroughsandrapid

qualityincrease

GoogleVeo3’srelease

showcasesaudio-video

breakthroughs,drivingadoption

andnewusecases

Chinacontinuesto

demonstrateleadershipin

languageandvideo

ModelsfromChineseAIlabs

occupytopspotsforopen

weightslanguagemodelsandon

thevideoleaderboard

PlayersintheAIvaluechaindifferinlevelsofverticalintegration;GooglecontinuestostandoutasthemostverticallyintegratedfromTPUacceleratorstoGemini

KeyPlayersintheAIValueChain(Non-Exhaustive)

No

presence

Strongpresence

Classificationsareindicativeanddeterminedbasedonarangeoffactorsincludingmarketshareandstrengthofoffering

Anthropic

Microsoft

DeepSeek

Snowflake

Databricks

SambaNova

Together.ai

Fireworks

DeepInfra

OpenAI

Google

Amazon

Alibaba

Perplexity

Cohere

Cerebras

Nebius

Meta

Mistral

NVIDIA

Groq

AMD

xAI

Applications

Foundation

models(firstparty)

Cloud Inference(firstparty)

AcceleratorHardware

8

Source:Companywebsite

BigtechnologycompaniesarecontinuingtoplayacrossallAImodalitieswhilesmallerchallengerstendtofocusonspecificmodalities

Keyplayerswithfirst-partymodelsbytypeofAINomodelExistingmodel

Language

Speech

Image

Video

Anthropic

Microsoft

OpenAI

GGoogle

Mistral

Amazon

NVIDIA

Meta

xAI

Adobe

ElevenLabs

Perplexity

Alibaba

Bytedance

Tencent

Baidu

DeepSeek

Kuaishou

MiniMax

Cohere

Midjourney

AI21Labs

AI21labs

9

Source:Companywebsite

01.IndustryOverview

AnumberofAILabsnowhavemodelsnearthefrontierofintelligence;xAIhastheleadingmodelwithGrok4,achievingthisfeatinlessthan500dayssincetheirfirstmodels'launch

FrontierLargeLanguageModel(LLM)Intelligence,OverTime

ArtificialAnalysisIntelligenceIndexv2(incorporatesMMLU-Pro,GPQADiamond,Humanity'sLastExam,LiveCodeBench,SciCode,AIME2024,MATH-500)

o3-Pro

Gemini2.5Pro

Grok4

DeepSeekR10528

Claude4

Opus(Extended

Thinking)

G

Llama4Maverick

?xAIleadstheintelligencefrontierforthefirsttime:xAIGrok4achievesthehighestintelligencescore(73)ontheArtificialAnalysisIndex,surpassingOpenAI'so3-pro(71),GoogleGemini2.5Pro(70),andDeepSeekR1(68)

?Open-sourcemodelsreachfrontierperformance:DeepSeekR1ranksamongthemostintelligentmodelsglobally,provingopen-weightsarchitecturescancompetewithproprietarysolutions

?OpenAI’sleadfaceschallenge:TheintelligencefrontierisnowfiercelycontestedbymultipleAIlabs,challengingOpenAI’slong-heldleadership

10

Source:ArtificialAnalysisindependentbenchmarking

02

02.LanguageModels

xAI,OpenAI,andGoogleleadfrontierintelligencewiththeirlatestreasoningmodels,followedcloselybyotherlabs

LeadingLargeLanguageModels(LLMs),byAIlab

Commentary

HighestArtificialAnalysisIntelligenceIndexv2achievedbyeachAILab

NON-EXHAUSTIVE

?OpenAIlosesfrontierforthefirsttime:xAI’sGrok4isnowthemostintelligentlanguagemodel,sittingaheadofo3-

pro,OpenAI’sfrontiermodel

?xAI,OpenAIandGoogleleadfrontierintelligence:Latestreasoningmodelsfromthreelabsholdthetop5positions

?Reasoningmodelscontinuetodominate:Q2‘25

continuestoseereasoning

modelssolidifytheirpositionastheclearestpathtohigherintelligenceindexscores

?Globalcompetition

intensifies:Labslike

DeepSeek,MiniMax,and

Alibabacontinuetoclosegap

12

Source:ArtificialAnalysisindependentbenchmarking

02.LanguageModels

Models:Overthepastyear,OpenAIhasmaintaineditslead,GoogleGeminiandDeepSeekhavesurged,andMetaLlamaandMistralhavefallen

DemandforTop10LLMFamiliesinMay2025

+1%

84%83%

+49%

80%

+21%

WhichLLMfamiliesareyouusingorconsideringusing?N=270(2024)and591(2025)

20252024

Changebetween2024&2025,p.p.

fromArtificial

AdoptionSurvey

H12025

67%

+53%

-6%

53%

49%

46%

43%

-15%

+31%

37%

31%

+17%

31%+25%

25%

+14%

21%

22%

14%

4%

0%DeepSeek

0%

PerplexitySonarMicrosoftPhi

0%0%

GoogleGemini

OpenAI(GPT/o)

AnthropicClaude

MetaLlama

xAIGrokAlibabaQwenMistral

13

Source:ArtificialAnalysisAIAdoptionSurvey–H12025

02.LanguageModels

OpenSource:Openweightslanguagemodelscontinueto

improve,thegaptoleadingproprietarymodelsstayedsimilar

LeadingLanguageModelsbyLicenseType,OverTime

Commentary

ArtificialAnalysisIntelligenceIndex(incorporatesMMLU-Pro,GPQA,Humanity'sLastExam,LiveCodeBench,SciCode,AIME,MATH-500)

o3-pro

Gemini2.5Pro

Grok4

DeepSeek-R10528

DeepSeekR1

?Openweightsclosethegaptoproprietary

models:Thereleaseof

DeepSeekR10528inMayfurtherreducedthe

intelligencegaptoleadingproprietarymodelsfrom

GoogleandOpenAI

(similartoreleaseof

DeepSeekR1);thereleaseofGrok4hassince

widenedthisgap

?Proprietaryandopen

weightsmodelscontinuetheirrapidrelease

cadence:Q2‘25

continuedtoseefrequentincrementalimprovementsdrivethefrontier

14

Source:ArtificialAnalysisindependentbenchmarking

02.LanguageModels

OpenSource:LeadingproprietarymodelsarefromUSlabs,whileChinaleadstheopenweightsintelligencefrontier

LeadingLanguageModelsbyLicenseType

Commentary

ArtificialAnalysisIntelligenceIndexv2(incorporatesMMLU-Pro,GPQA,Humanity'sLastExam,LiveCodeBench,SciCode,AIME,MATH-500)

?Proprietarymodels

continuetoleadfrontierintelligence:Proprietary

reasoningmodelsfromUSlabsleadinoverall

intelligence

?Chinademonstratesopenweightsleadership:

NON-EXHAUSTIVE

Leadingopenweights

modelsarefromChineseAIlabs(DeepSeek,MiniMax,

Alibaba,Moonshot)

?Proprietarymodels

marginallyleadfornon-

reasoningmodels:Claude

4Opusiscurrentlythe

mostintelligentnon-

reasoningmodel,followedcloselybyKimiK2

15

Source:ArtificialAnalysisindependentbenchmarking

02.LanguageModels

Countryview:ModelsfromlabsintheUSandChinacontinuetodominatetheintelligencefrontier

LeadingLanguageModelsbyCountryofOrigin

ArtificialAnalysisIntelligenceIndexv2(incorporatesMMLU-Pro,GPQA,Humanity'sLastExam,LiveCodeBench,SciCode,AIME,MATH-500)Commentary

?USmaintainsleadershipin

frontierreasoning:US-basedlabscontinuetoholdthetop

spotsontheIntelligenceIndexwiththeirpremierreasoning

modelslikeGrok4,o3-proandGemini2.5Pro

?Q2sawlimiteddisruption

fromothercountries.Francemaintainsapresencewith

MagistralMedium,while

UpstageAI’sSolarPro2modelbroughtSouthKoreatothe

frontierforthefirsttime

?Overall,theglobalfrontier

remainshighlyconcentrated,withtheUSandChina

continuingtodefinethepaceanddirectionofcutting-edgemodeldevelopment

16

Source:ArtificialAnalysisindependentbenchmarking

17

02.LanguageModels

…computedemandcontinuestoincrease

Whileefficiencygainshavebeenmade…

DeepDivenext

Newapplicationscontinuetodemandmorecompute:asingledeepresearchquerycancost>10xanoriginalGPT-4query

GPT-4levelintelligenceisnow100xcheaperthanoriginalGPT-4

A.SmallerModels

B.SoftwareEfficiency

Algorithmicandtrainingdataimprovements

C.HardwareEfficiency

Nextgenerationacceleratorsoffer

haveallowedsmaller

Inferenceoptimizations(e.g.FlashAttention)

modelstogetsmarter

improveefficiency

~1/3x

compute

~1/10x

compute

morecompute

efficiency

~1/3x

costs

~20x

requests/use

F.AIAgents

~10x

Agentschainmultiple

tokens/query

~5x

requeststoLLMsto

completetasks

autonomously

E.ReasoningModels

compute/query

D.LargerModels

Scalinglawscontinue

todemandhigher

parametercountsfor

greaterintelligence

Significantincreasein

outputtokenswhen

Figuresarehighlyindicativeandservetoillustratethedirectional

impactofeachfactorimpactingcost

models‘think’before

answering

02.LanguageModels

B.SoftwareEfficiency:EfficientmodelscombinedwithnewacceleratorskeptslashingAI

LanguageModelInferencePricingbyIntelligenceClass,OverTime

PriceinUSDper1milliontokens(blendedinputtooutputtokenprice3:1);ArtificialAnalysisIntelligenceIndexv2(incorporates7evaluations)

Commentary

inferencecoststhroughoutQ2

GPT-4

GPT-4o

NON-EXHAUSTIVE

?

Q22025acceleratesthe

slideininferencecost:fromApriltoJune,pricesfell

acrosseveryintelligence

bandasDeepSeekR10528,Qwen38B,andGemma3nE4BInstructslashedcosts

whileliftingscores

GPT-3.5Turbo

o1-mini

DeepSeekR1DistillLlama8B

Gemini2.0FlashLite

DeepSeek-R10528

Qwen38B

Gemma3n

E4BInstruct

/

?

CapableAIisbecoming

moreaccessibleand

commoditized:duringQ2

2025,thepriceoffrontier-

levelinference(Intelligence

Index≥50)droppedbynearly75%,slidingfrom$0.26to

just$0.063permilliontokens

18

Source:ArtificialAnalysisindependentbenchmarking

02.LanguageModels

B.SoftwareEfficiency:ThroughputsignificantlyincreasedinQ22025acrossmodel

classes,butend-userwaittimesaresometimesgrowingduetolongreasoningchains

LanguageModelOutputSpeedbyIntelligence,OverTime

Totaloutputtokenspersecond,ArtificialAnalysisIntelligenceIndexv2(incorporates7leadingevaluations)

?AQ22025speedsurge

overcamethetrade-off

againstintelligence:a

significantleapin

inferenceperformance

occurredinthesecond

quarterof2025,andnewreleasesmadehighly-

intelligentmodels(Index

>=50)thefastestcategoryforthefirsttime

?Latencyparadox:despitehigherthroughput,end-to-endusecanbeslowerasreasoningandagentic

tasksgeneratetensof

thousandsoftokensandchainmultiplecalls,fullyoffsettingspeedgains

Commentary

NON-EXHAUSTIVE

Gemini2.5Flash-Lite(Reasoning)——

Gemini2.5Flash-Lite

NovaMicro——Gemini1.5Flash-8B

19

Source:ArtificialAnalysisindependentbenchmarking

02.LanguageModels

E.ReasoningModels:Reasoningcoststimeandcompute:reasoningmodelsuseupto

10xmoretokenstorespondtothesamepromptsasnon-reasoningmodels

OutputTokensUsedtoRunArtificialAnalysisIntelligenceIndex

ArtificialAnalysisIntelligenceIndexv2(incorporates7leadingevaluations),OutputTokensUsedinArtificialAnalysisIntelligenceIndex(~5Minputtokens)Reasoningmodels

NON-EXHAUSTIVE

~78M1

Avg.totaloutputtokensfor

reasoningmodels

~10M1

Avg.totaloutputtokensfornon-

reasoningmodels

20

1.BasedonrepresentativemodelsincludedinthechartSource:ArtificialAnalysisindependentbenchmarking

21

02.LanguageModels

InQ22025wesawincreaseduseofagenticworkflowsandexplosivegrowthincodingagents,bothenabledbyaconnectionecosystemandnewmodeltrainingapproaches

KeyThemesinQ2‘25

Applicationsmovetowards‘a(chǎn)genticbydefault,

?

?

AgenticworkflowscontinuetobecomeembeddedinawiderangeofAIapplicationsthatpreviouslyusedlinearexecutionandminimaltooluse,suchaschatbots,terminals,anddataanalysistools

DeepresearchagentsbecametablestakesformajorchatbotsandsomesmallerChineselabentrants

Ecosystemofconnections

?

ApplicationssuchasChatGPTandClaudeexpandedtheirsuiteofintegrations,bothwithinternallydevelopedtools

continuestogrowand

enablenewfunctionality

?

andincreasingModelContextProtocol(MCP)compatibility

FirstandthirdpartyMCPserversproliferatedacrossarangeofAPIsandbusinesses,withstrongusageindeveloperandconsumerapplications

Codingagentsseerapidgrowth

?

?

Q2sawanunprecedentedvolumeofnewcodingagentproducts,with12majorcodingagentlauncheswithinthequarter,includingfrontierlabproductslikeOpenAICodexandGeminiCLI

Codingagentusagehasrapidlygrown;approximatelyhalfoftherespondentstotheArtificialAnalysisAIAdoptionSurveyuseorareconsideringusingCursor

AgenticmodelusewilldriveupLLMusagecosts

?

Agentsincuradditionalusageoftokensandtools,drivingincreasedcosts;deepresearchAPIslaunchedinQ2havedemonstratedcostsofupto$28forasinglecomplexqueryintesting

Trainingfocusesonagentsandlong-horizontooluse

?

?

Reasoningmodelsandreinforcementlearninghasenabledmoreeffectivetooluse,includinginterleavedwithmodelthinkingbeforeproducingresponsestousers

Modelcreatorsincreasedtheemphasisintrainingonlong-runningtasksandagenticworkflowsformodelssuchastheClaude4familyandKimiK2

02.LanguageModels

AgentsareautonomoussystemsdrivenbyLLMs…

Whatareagents?

“SystemswhereLLMsdynamicallydirecttheirownprocessesandtoolusage,maintainingcontroloverhowtheyaccomplishtasks”

“Agentsrepresentsystemsthatintelligentlyaccomplishtasks,rangingfromexecutingsimpleworkflowsto

pursuingcomplex,open-endedobjectives”

“AIagentsareautonomoussystemspoweredbylargelanguagemodels(LLMs)that,givenhigh-level

instructions,canplan,usetools,carryoutstepsof

processing,andtakeactionstoachievespecificgoals”

AIagentsareLLM-drivensystemsthatact

autonomouslyandusetoolstocomplete

tasksend-to-end

actions

1.CanincludedirectvisioncapabilityorprocessingofotherdatasuchaswebsiteHTMLtoidentifySource:Companywebsite;Anthropic‘Buildingeffectiveagents’Agentworkflowdefinition

Fundamentally,agentsineverydomainruninaloopandtake

actionsbyusingtools,suchassearchingtheweborwritingtoafile

User

Agentdecideswhen

thetaskis‘complete’

Usersmakeinitialrequests,andthe

agentmayengagetheminfurther

turnswhereneeded(e.g.,toclarify)

Agent

Completetask

LargeLanguageModel

Toolsetandenvironment(exampletoolinclusions)

Filesystemaccess

APIintegrations

Codeexecutionenvironments

MCPservers

02.LanguageModels

Builtontheimprovingintelligenceoflanguagemodels,AIagentsofferkeybenefitscomparedtotraditionalworkflowsandareseeingsuccessinseveraldomains

OverviewofagentbenefitsKeydomainsshowingearlysuccesses

Finds,interprets,editsandtestssourcecodetocompletesoftwareengineeringtasks

Coding

AgenticapproachesenablenewAI-basedapplicationsduetoarangeofkeybenefitscomparedtostaticworkflows:

Deep

research

Parses(andpotentiallyclarifies)aresearchqueryandlaunchesachainoftargetedresearchquerieswhile

controllingitsresearchflowtosynthesizeananswer

Computeruse

Interpretsusercommands,‘looks’atadesktopor

browserwindow1,andautonomouslychainsclicks,keystrokes,shellcommands,andAPIcallsto

completearbitrarytask

Customersupport

Customeragentinlivespeechortextchatwhich

identifiesintentandrespondstocustomersinreal

time,whilechainingrequiredapp,CRMorAPIcallstocompletethetask(orhandofftoahumanagent)

1.Dynamicplanning,tasktracking,andexecutionfor

complexunknowntaskrequirementstopursuewell-definedgoals

2.Integrationwithawiderangeofsystemsandprocesses

acrossadomainwithoutaclearsequenceofdependenciesor‘chains’ofusetocompletetasks

3.Naturalcollaborationtocompletetasks,includingengaginghumanusersinthelooptoclarifyorcontinuetasks,or

coordinatingwithotheragentswithadditionalcapabilities

SalesIdentifiespotentialleads,executespersonalized

outreach,andintegrateswithsalestools

4.Gracefulerrorrecoveryfromfeedbackwhereerrorsoccur,evenwithuniqueorunexpectedfailuremodes

ForadeepdiveonthelatestprogressinAIagents,seetheArtificialAnalysisQ2AgentsandApplicationsReport

02.LanguageModels

Severalcompetingplayersareemerginginthebigagentdomainsin

2025;leadinglabsarefocusedoncoding,research,andcomputeruse

NON-EXHAUSTIVE

Frontierlabproduct

Domain

Illustrativeproducts&players

ACoding

母GitHubcopilot

I-replit

BDeepresearch

上?MistralAl

CComputeruse

ADEPT

Gcomet&brouseruse

D

Customersupport

漲Fin

?freshworks診Decagon

E

Sales

M11x

ppersanaAl

AisDR

?RelevanceAl

Thesedomainsareshowingthemostprogressin

commercialoff-the-shelfproductsandresearch

previews,whileotherusecasesarefollowing.

Inparallel,arangeofprovidersareenablinguserstobuild

customAIagentsfortheirusecases:

?RelevanceAl

stackAl

02.LanguageModels

A.AICodingTools:GitHubCopilotandCursordominatethemarketasthemostpopular

AIcodingtools,withasignificantleadoverClaudeCodeandGeminiCodeAssist

DemandforCodingTools

WhichAItoolsareyouusingorconsideringusingthisyear?N=955

ExcerptfromArtificial

AnalysisAIAdoptionSurvey

ReportH12025

25

Source:ArtificialAnalysisAIAdoptionSurvey–H12025

26

02.LanguageModels

A.Codingagents:Codingagentlauncheshaveacceleratedin2025,withalargefocusoncommand-lineinterfaces

NON-EXHAUSTIVE

Majorcodingagentproductlaunchesbyquarter

Numberofcodingagentlaunchesidentified,basedonprimaryformfactor1

IDEextensionCloudcodingagent

DedicatedIDEAppbuilder

Localnon-IDE(incl.CLIs)

12

2

3

NotablereleasesinQ2included:?GitHubCopilotCodingAgent?CursorBackgroundAgents

?OpenAICodexCLIandcloud?GoogleGeminiCLIandJules

InJuly2025,AmazonlaunchedKiro,anAIIDEinpublicpreview

6

5

1

1

2

2

5

3

1

2

3

1

1

1

1

2

1

1

1

1

1

1

1

1

1

0

1

2

2023-Q12023-Q22023-Q32023-Q42024-Q12024-Q22024-Q32024-Q42025-Q12025-Q2

Releasesinclude

yz

replit

windsurf的augmentcode

ROOCOOE

OpenAI

Codex

Backgroundagents

?GitHubcopilot

1.Primaryformfactorisassessedqualitativelywhenmultiplemodesapply

Note:Releasetimingsarebestestimates,basedongeneralorpublicavailabilitywherepossible.Whererelevant,timingisbasedonwhenAIcodingagentsbecameacoreproductcapability

03ImageandVideoModels

StateofAI–Q22025

03.ImageandVideoModels

Q2‘25sawashiftinprogresstoVideomodels,withaudiosupportandbreakthroughsinquality,whileopenweightsmodelprogressslowedinbothimageandvideo

KeyThemesinQ2‘25

Videomodelsbegintosupportaudio

?

?

Veo3releasedinMay2025becomesthefirsthighquality,mainstreammodelthatnativelysupportedaudiogenerationaspartofavideomodel,drivingstrongadoption

Veo3,sdifferentiatedaudiosupportgivesitstrongpricingpowerat$0.75/sof720pvideowithaudio,surpassingcomparablemodelssuchasSeedance1.0at~$0.13/sof1080pvideo,andHailuo2at~$0.08/sof1080pvideo

Videomodelsseecontinuedbreakthroughsinquality

?

?

?

Videomodelsseeabreakthroughinquality,withSeedance1.0overtakingbothQ1leaders:Veo2intexttovideoby~150ELO,andKling1.6Proinimagetovideoby~200ELOpoints

Labsshiftfocustoimagetovideogenerations,withalargerELOjumpthantexttovideo,andmodelssuchasMidjourneyV1andKling2.1Proavailableonlyinimagetovideovariants

Openweightsvideomodelslagbehindproprietaryalternatives,withAlibabaWan2.1stillrepresentingtheSOTAforopenweightstexttovideogenerationandLTXVideov0.9.713Branking16thoverallforimagetovideoontheArtificialAnalysisleaderboard

Imageeditingmodelslaunched

?

?

Instructionbasedimageeditingmodelsbecomepopular,withGPT-4ocontinuingtoholdthelead,butFLUX.1Kontext[max]andHiDream-E1.1launchedascompetitivemodelsinQ2

Openweightsima