Reflection vs Anthropic vs Google vs DeepSeek

Head to head on 21 boards: Anthropic places best on 14, Google places best on 6, DeepSeek places best on 1 and Reflection places best on 0. All 4 ranked on 2 of 28 boards.

Two at a time on this screen. Tap one to swap it in.

RReflectionAnthropicGoogleDeepSeek

Board	Read	RReflection	Anthropic	Google	DeepSeek
Chat2
Text ArenaArena ratingtodaytoday	Oct 21todaytoday		#21,507Claude Opus 5.5	#11,525best of the pickedGemini 4 Argon	#341,475DeepSeek V4.1 Flash
MultiChallengeScoretodaytoday	Oct 21todaytoday		#667.7%Claude Haiku 5.5	#471.4%best of the pickedGemini 3.1 Pro Preview	#2646.1%DeepSeek V3P1
Reasoning7
ARC-AGI-2Scoretodaytoday	Oct 21todaytoday	#4623.5%Beam 1.0	#393.3%best of the pickedClaude Opus 5.5	#591.3%Gemini 4 Argon	#2172.9%DeepSeek V4.1 Flash
GPQA DiamondAccuracytodaytoday	Oct 21todaytoday		#395.6%Claude Sonnet 5.5	#196.4%best of the pickedGemini 4 Argon	#2591.7%DeepSeek V4 Pro 0813
Epoch ECIECItodaytoday	Oct 21todaytoday		#1168.8best of the pickedClaude Opus 5.5	#5166.4Gemini 4 Argon	#32155.9DeepSeek V4.1 Flash
LiveBenchGlobal averagetodaytoday	Oct 21todaytoday		#183.4%best of the pickedClaude Fable 5.1	#1578.8%Gemini 3.7 Flash	#881.1%DeepSeek V4.1 Flash
MMLU-ProAccuracy · staletodaytoday	Oct 21todaytoday		#489.1%Claude Opus 4.6	#191.2%best of the pickedGemini 3.1 Pro Preview	#3385.0%DeepSeek-V3.2
AA IntelligenceIndextodaytoday	Oct 21todaytoday		#158.0best of the pickedClaude Opus 5.5	#453.0Gemini 4 Argon
HLEAccuracytodaytoday	Oct 21todaytoday		#246.5%best of the pickedClaude Fable 5.1	#346.4%Gemini 3.1 Pro Preview
Coding4
WebDev ArenaArena ratingtodaytoday	Oct 21todaytoday		#11,814best of the pickedClaude Opus 5.5	#81,678Gemini 4 Argon	#191,619DeepSeek V4.1 Flash
SWE-bench VerifiedResolved · staletodaytoday	Oct 21todaytoday		#176.8%best of the pickedClaude Opus 4.5	#275.8%Gemini 3 Flash	#1170.0%DeepSeek-V3.2
Aider PolyglotPass rate · staletodaytoday	Oct 21todaytoday		#1072.0%Claude Opus 4	#383.1%best of the pickedGemini 2.5 Pro (Jun 2025)	#874.2%DeepSeek-V3.2
SWE-Bench ProResolvedtodaytoday	Oct 21todaytoday		#198.0%best of the pickedClaude Opus 5	#392.2%Gemini 4 Argon
Agents and tools3
BFCLOverall accuracy · staletodaytoday	Oct 21todaytoday		#177.5%best of the pickedClaude Opus 4.5	#372.5%Gemini 3 Pro	#1256.7%DeepSeek-V3.2
MCP AtlasPass ratetodaytoday	Oct 21todaytoday		#287.2%best of the pickedClaude Fable 5.1	#683.6%Gemini 3.5 Flash
Search ArenaArena ratingtodaytoday	Oct 21todaytoday		#21,253best of the pickedClaude Opus 4.6	#91,210Gemini 3.1 Pro Grounding
Vision2
Vision ArenaArena ratingtodaytoday	Oct 21todaytoday		#11,308best of the pickedClaude Fable 5	#51,297Gemini 3.7 Flash	#391,260DeepSeek V4.1 Flash
Document ArenaArena ratingtodaytoday	Oct 21todaytoday		#11,516best of the pickedClaude Opus 5	#161,463Gemini 3.5 Flash
Image2
Text-to-ImageArena ratingtodaytoday	Oct 21todaytoday			#61,328Gemini Nano Banana 2.1
Image EditArena ratingtodaytoday	Oct 21todaytoday			#61,428Gemini Nano Banana 2.1
Video2
Text-to-VideoArena ratingtodaytoday	Oct 21todaytoday			#11,516Gemini Omni 1.1 Flash
Image-to-VideoArena ratingtodaytoday	Oct 21todaytoday			#21,488Gemini Omni 1.1 Flash
Speech2
Speech ArenaArena ratingtodaytoday	Oct 21todaytoday			#51,278Gemini 3.8 Flash TTS
Open ASRAverage WERtodaytoday	Oct 21todaytoday			#5911.65 WERgemma-4-E2B-it (google)
Embeddings1
MTEBMean task scoretodaytoday	Oct 21todaytoday			#568.4%gemini-embedding-001 (google)
Speed1
AA SpeedOutput speedtodaytoday	Oct 21todaytoday	#6345 tok/sBeam 1.0	#12236 tok/sClaude Haiku 5.5	#4454 tok/sbest of the pickedGemini 3.5 Flash-Lite	#16209 tok/sDeepSeek V4.1 Flash
Price1
AA PriceBlended pricetodaytoday	Oct 21todaytoday		#3$0.20best of the pickedClaude Haiku 5.5	#13$1.50Gemini 3.8 Flash	#8$0.53DeepSeek V4.1 Flash
Usage1
OpenRouterWeekly tokenstodaytoday	Oct 21todaytoday		#56.8T tokClaude Opus 5.5		#127.3T tokbest of the pickedDeepSeek V4.1 Flash

Green marks the best placing among the picked where two or more are ranked; ties are marked for each. Ranks are each board's own order (lower is better), scores as the board published them. Dash: not on that board.