Skip to main content

Head-to-head comparison

GPT-5.4 vs. DeepSeek-V4-Flash

Compare published benchmark results from matching versions and cohorts, API pricing, and technical specifications. Any documented test setup differences remain visible.

shared result series
60
distinct benchmarks
2
latest retrieval
2026-09-29

Change model selection

Specifications

General specifications

Release date, availability, context window, architecture, and published parameter counts at a glance. Undisclosed details remain marked as unavailable.

Specifications for GPT-5.4 and DeepSeek-V4-Flash
AttributeGPT-5.4DeepSeek-V4-Flash
ProviderOpenAIDeepSeek
ReleasedMar 5, 2026Apr 24, 2026
AvailabilityActiveLimited access
Model typeProprietaryOpen Weights
ParametersNot published284B, 13B active
ArchitectureNot publishedMixture of Experts
Context window1,050,000 tokens1,000,000 tokens
Knowledge cutoffAugust 31, 2025Not published
Technical sourcesModelModel

Pricing

API pricing compared

The table compares official standard API rates per 1M tokens, including published cache pricing. Batch, fast, priority, regional surcharges, tool calls, and cloud platform pricing are excluded.

API pricing for GPT-5.4 and DeepSeek-V4-Flash
Price typeGPT-5.4DeepSeek-V4-Flash
API input$2.5 ≤272K / $5 >272K$0.05
API output$15 ≤272K / $22.5 >272K$0.65
Cache read$0.25 ≤272K / $0.5 >272KNot listed
Cache write (5 min.)Not listedNot listed
Cache write (1 hr.)Not listedNot listed
Price verified09/22/2026 Source10/03/2026 Source
GPT-5.4DeepSeek-V4-Flash
API inputUSD per 1M tokens, base tier
GPT-5.4$2.5
DeepSeek-V4-FlashLower price$0.05
API outputUSD per 1M tokens, base tier
GPT-5.4$15
DeepSeek-V4-FlashLower price$0.65

Key to the marks:better valuebetter, but not conclusivetie

Pricing by provider

Direct developer pricing and provider endpoint tariffs routed through OpenRouter are listed separately. Prices are in USD per 1M tokens. Context-dependent direct tiers plus regional, Flex, Priority, and other OpenRouter tariffs remain individually identifiable.

Provider pricing for GPT-5.4 and DeepSeek-V4-Flash
ModelProviderPurchase route and tariffInput per 1M tokensOutput per 1M tokensSource
GPT-5.4OpenAIDirect from developerUp to 272,000 context tokens$2.5$15Source09/22/2026
OpenAIDirect from developerAbove 272,000 context tokens$5$22.5Source09/22/2026
Amazon BedrockVia OpenRouteramazon-bedrock/us-east-1$2.75$16.5Source10/03/2026
AzureVia OpenRouterazure$2.5$15Source10/03/2026
AzureVia OpenRouterazure/eu$2.75$16.5Source10/03/2026
AzureVia OpenRouterazure/us$2.75$16.5Source10/03/2026
OpenAIVia OpenRouteropenai$2.5$15Source10/03/2026
OpenAIVia OpenRouteropenai/fast$5$30Source10/03/2026
OpenAIVia OpenRouteropenai/flexLowest listed price: $1.25Lowest listed price: $7.5Source10/03/2026
DeepSeek-V4-FlashAlibabaVia OpenRouteralibaba$0.176$0.528Source10/03/2026
AtlasCloudVia OpenRouteratlas-cloud/fp4$0.44$1.32Source10/03/2026
BaiduVia OpenRouterbaidu/fp8$0.44$1.32Source10/03/2026
BaseTenVia OpenRouterbaseten/fp8#1$0.13$0.26Source10/03/2026
BaseTenVia OpenRouterbaseten/fp8#2$0.13$0.26Source10/03/2026
CloudflareVia OpenRoutercloudflare$0.44$1.32Source10/03/2026
CohereVia OpenRoutercohere$0.14$0.28Source10/03/2026
CoreWeaveVia OpenRoutercoreweave/fp8$0.13$0.28Source10/03/2026
DeepInfraVia OpenRouterdeepinfra/fp8$0.06$0.18Source10/03/2026
DigitalOceanVia OpenRouterdigitalocean$0.119$0.238Source10/03/2026
GMICloudVia OpenRoutergmicloud/fp8$0.286$0.858Source10/03/2026
InceptronVia OpenRouterinceptron/fp4$0.05$0.65Source10/03/2026
Mancer 2Via OpenRoutermancer/fp8$0.2$0.6Source10/03/2026
MorphVia OpenRoutermorph/bf16$0.142$0.4Source10/03/2026
NextBitVia OpenRouternextbit/fp8$0.352$1.056Source10/03/2026
NovitaVia OpenRouternovita/fp8$0.4092$1.2276Source10/03/2026
OpenInferenceVia OpenRouteropen-inference/fp8Lowest listed price: $0.0128$0.6026Source10/03/2026
ParasailVia OpenRouterparasail/fp8$0.14$0.28Source10/03/2026
PhalaVia OpenRouterphala$0.308$0.924Source10/03/2026
RekaVia OpenRouterreka$0.021$0.528Source10/03/2026
RelaceVia OpenRouterrelace/fp4$0.0171$1.28Source10/03/2026
Sail ResearchVia OpenRoutersail-research/fp4$0.019$0.3Source10/03/2026
Sail ResearchVia OpenRoutersail-research/us$0.019$0.42Source10/03/2026
SiliconFlowVia OpenRoutersiliconflow/fp8$0.22$0.66Source10/03/2026
StreamLakeVia OpenRouterstreamlake/fp8$0.044Lowest listed price: $0.132Source10/03/2026
TogetherVia OpenRoutertogether$0.14$0.28Source10/03/2026
VeniceVia OpenRoutervenice$0.175$0.35Source10/03/2026
WaferVia OpenRouterwafer/fast$0.12$0.7Source10/03/2026

Performance

Shared benchmarks

Only published results with the same benchmark version, task, metric, and comparison cohort are paired. Different reasoning levels, agents, harnesses, or output limits appear directly in the table.

Capability profile from matched benchmarks

Each axis averages directly matched benchmark families. A value of 100 means the stronger value within this pair, not a universal quality score.

255075100CodingReasoningMathematicsKnowledgeFinanceLegalCybersecurityHealth
GPT-5.4
DeepSeek-V4-Flash
Coding: 100 / 97.8 (1 benchmark family)
Reasoning: 100 / 97.7 (1 benchmark family)
Mathematics: 100 / 97 (1 benchmark family)
Knowledge: 100 / 97.9 (1 benchmark family)
Finance: 100 / 97.2 (1 benchmark family)
Legal: 100 / 98 (1 benchmark family)
Cybersecurity: 100 / 91.9 (1 benchmark family)
Health: 100 / 99.3 (1 benchmark family)

60 matched rows from 2 benchmark families were considered. The radar shows 8 of 9 comparable categories.

GPT-5.4DeepSeek-V4-Flash
CyberBench v1.1Patch. Higher is better. No documented setup difference
GPT-5.4Winner81.356 %
DeepSeek-V4-Flash72.414 %
LMArena Text, Style ControlOverall. Higher is better. No documented setup difference
GPT-5.4Winner1,464.964
DeepSeek-V4-Flash1,436.082
CyberBench v1.1PoC. Higher is better. No documented setup difference
GPT-5.4Winner64.407 %
DeepSeek-V4-Flash61.017 %
LMArena Text, Style ControlOverall. Higher is better. No documented setup difference
GPT-5.4Winner1,475.198
DeepSeek-V4-Flash1,438.511
LMArena Text, Style ControlBusiness, management and finance. Higher is better. No documented setup difference
GPT-5.4Winner1,474.623
DeepSeek-V4-Flash1,438.526
LMArena Text, Style ControlBusiness, management and finance. Higher is better. No documented setup difference
GPT-5.4Winner1,481.559
DeepSeek-V4-Flash1,434.189
LMArena Text, Style ControlChinese. Higher is better. No documented setup difference
GPT-5.4Winner1,493.153
DeepSeek-V4-Flash1,471.267
LMArena Text, Style ControlChinese. Higher is better. No documented setup difference
GPT-5.4Winner1,506.812
DeepSeek-V4-Flash1,479.924
LMArena Text, Style ControlCoding. Higher is better. No documented setup difference
GPT-5.4Winner1,511.177
DeepSeek-V4-Flash1,484.094
LMArena Text, Style ControlCoding. Higher is better. No documented setup difference
GPT-5.4Winner1,520.571
DeepSeek-V4-Flash1,479.555
LMArena Text, Style ControlCreative writing. Higher is better. No documented setup difference
GPT-5.4Winner1,434.479
DeepSeek-V4-Flash1,407.791
LMArena Text, Style ControlCreative writing. Higher is better. No documented setup difference
GPT-5.4Winner1,442.663
DeepSeek-V4-Flash1,410.703

Key to the marks:better valuebetter, but not conclusivetie

Benchmark scores for GPT-5.4 and DeepSeek-V4-Flash
Benchmark and taskGPT-5.4DeepSeek-V4-FlashSource
CyberBench v1.1Patch
Test details
No documented difference in the test setupVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05
Winner81.356 %
72.414 %
Vals AI
Source details
Data as of: 08/03/2026
LMArena Text, Style ControlOverall
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 67,632 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 51,871 anonymous pairwise votes.
Winner1,464.964
95% confidence interval: ±3.702The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,436.082
95% confidence interval: ±4.048The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
CyberBench v1.1PoC
Test details
No documented difference in the test setupVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05
Winner64.407 %
61.017 %
Vals AI
Source details
Data as of: 08/03/2026
LMArena Text, Style ControlOverall
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 64,479 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 51,747 anonymous pairwise votes.
Winner1,475.198
95% confidence interval: ±3.778The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,438.511
95% confidence interval: ±4.115The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlBusiness, management and finance
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 13,882 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 10,436 anonymous pairwise votes.
Winner1,474.623
95% confidence interval: ±6.376The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,438.526
95% confidence interval: ±6.971The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlBusiness, management and finance
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 13,092 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 10,641 anonymous pairwise votes.
Winner1,481.559
95% confidence interval: ±6.485The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,434.189
95% confidence interval: ±6.982The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlChinese
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 4,364 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 2,996 anonymous pairwise votes.
Winner1,493.153
95% confidence interval: ±9.833The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,471.267
95% confidence interval: ±11.424The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlChinese
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 4,159 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 2,992 anonymous pairwise votes.
Winner1,506.812
95% confidence interval: ±10.001The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,479.924
95% confidence interval: ±11.42The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlCoding
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 18,919 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 15,262 anonymous pairwise votes.
Winner1,511.177
95% confidence interval: ±5.805The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,484.094
95% confidence interval: ±6.23The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlCoding
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 17,577 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 15,239 anonymous pairwise votes.
Winner1,520.571
95% confidence interval: ±5.854The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,479.555
95% confidence interval: ±6.265The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlCreative writing
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 11,135 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 8,539 anonymous pairwise votes.
Winner1,434.479
95% confidence interval: ±6.844The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,407.791
95% confidence interval: ±7.543The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlCreative writing
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 10,894 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 8,638 anonymous pairwise votes.
Winner1,442.663
95% confidence interval: ±6.988The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,410.703
95% confidence interval: ±7.575The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlEnglish
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 30,587 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 23,008 anonymous pairwise votes.
Winner1,468.151
95% confidence interval: ±4.916The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,444.92
95% confidence interval: ±5.29The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlEnglish
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 28,991 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 22,908 anonymous pairwise votes.
Winner1,476.387
95% confidence interval: ±4.966The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,449.94
95% confidence interval: ±5.323The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlEntertainment, sports and media
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 14,616 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 11,179 anonymous pairwise votes.
Winner1,432.028
95% confidence interval: ±6.282The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,404.567
95% confidence interval: ±6.937The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlEntertainment, sports and media
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 13,917 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 11,428 anonymous pairwise votes.
Winner1,441.47
95% confidence interval: ±6.441The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,408.063
95% confidence interval: ±6.896The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlExcluding ties
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 51,980 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 39,226 anonymous pairwise votes.
Winner1,471.069
95% confidence interval: ±4.836The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,431.809
95% confidence interval: ±5.308The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlExcluding ties
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 49,393 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 39,391 anonymous pairwise votes.
Winner1,484.279
95% confidence interval: ±4.933The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,436.008
95% confidence interval: ±5.353The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlExpert prompts
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 6,707 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 5,392 anonymous pairwise votes.
Winner1,492.198
95% confidence interval: ±8.255The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,457.264
95% confidence interval: ±8.778The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlExpert prompts
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 6,334 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 5,311 anonymous pairwise votes.
Winner1,514.737
95% confidence interval: ±8.393The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,466.306
95% confidence interval: ±8.867The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlFrench
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 2,420 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 2,006 anonymous pairwise votes.
Winner1,485.522
95% confidence interval: ±14.25The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,448.483
95% confidence interval: ±15.039The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlFrench
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 2,403 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 1,966 anonymous pairwise votes.
Winner1,499.866
95% confidence interval: ±14.273The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,459.263
95% confidence interval: ±15.286The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlGerman
Test details
No documented difference in the test setupNo clear leadVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 1,001 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 843 anonymous pairwise votes.
No clear lead1,449.588
95% confidence interval: ±19.823The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,424.769
95% confidence interval: ±21.221The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlGerman
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 1,092 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 906 anonymous pairwise votes.
Winner1,476.158
95% confidence interval: ±19.045The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,433.595
95% confidence interval: ±20.539The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlHard prompts
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 44,698 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 34,753 anonymous pairwise votes.
Winner1,487.189
95% confidence interval: ±4.475The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,458.919
95% confidence interval: ±4.767The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlHard prompts
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 42,103 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 34,700 anonymous pairwise votes.
Winner1,496.664
95% confidence interval: ±4.521The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,458.083
95% confidence interval: ±4.832The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlHard prompts, English
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 21,002 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 16,176 anonymous pairwise votes.
Winner1,486.858
95% confidence interval: ±5.628The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,463.433
95% confidence interval: ±6.065The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlHard prompts, English
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 19,649 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 16,039 anonymous pairwise votes.
Winner1,493.158
95% confidence interval: ±5.688The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,464.61
95% confidence interval: ±6.133The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlInstruction following
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 23,384 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 17,961 anonymous pairwise votes.
Winner1,460.276
95% confidence interval: ±5.39The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,428.187
95% confidence interval: ±5.811The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlInstruction following
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 21,896 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 18,051 anonymous pairwise votes.
Winner1,471.171
95% confidence interval: ±5.474The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,433.413
95% confidence interval: ±5.881The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlJapanese
Test details
No documented difference in the test setupNo clear leadVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 681 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 647 anonymous pairwise votes.
No clear lead1,445.374
95% confidence interval: ±24.201The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,398.363
95% confidence interval: ±24.906The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlJapanese
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 671 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 635 anonymous pairwise votes.
Winner1,479.239
95% confidence interval: ±24.937The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,424.242
95% confidence interval: ±25.166The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlKorean
Test details
No documented difference in the test setupNo clear leadVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 1,161 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 898 anonymous pairwise votes.
No clear lead1,425.994
95% confidence interval: ±19.269The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,392.925
95% confidence interval: ±20.977The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlKorean
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 1,107 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 951 anonymous pairwise votes.
Winner1,433.931
95% confidence interval: ±19.893The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,378.153
95% confidence interval: ±20.872The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlLegal and government
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 5,387 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 4,139 anonymous pairwise votes.
Winner1,475.748
95% confidence interval: ±9.028The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,447.478
95% confidence interval: ±9.874The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlLegal and government
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 5,240 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 4,107 anonymous pairwise votes.
Winner1,489.844
95% confidence interval: ±9.116The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,457.491
95% confidence interval: ±9.927The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlLife, physical and social science
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 11,096 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 8,423 anonymous pairwise votes.
Winner1,478.774
95% confidence interval: ±6.725The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,456.345
95% confidence interval: ±7.388The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlLife, physical and social science
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 10,659 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 8,509 anonymous pairwise votes.
Winner1,487.593
95% confidence interval: ±6.842The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,459.724
95% confidence interval: ±7.418The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlLonger queries
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 29,890 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 23,359 anonymous pairwise votes.
Winner1,477.734
95% confidence interval: ±5.219The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,447.83
95% confidence interval: ±5.578The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlLonger queries
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 28,372 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 23,501 anonymous pairwise votes.
Winner1,481.678
95% confidence interval: ±5.295The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,448.636
95% confidence interval: ±5.577The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlMath
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 3,551 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 2,688 anonymous pairwise votes.
Winner1,459.714
95% confidence interval: ±10.346The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,425.467
95% confidence interval: ±11.801The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlMath
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 3,387 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 2,619 anonymous pairwise votes.
Winner1,491.354
95% confidence interval: ±10.6The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,441.412
95% confidence interval: ±11.893The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlMathematical professions
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 3,730 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 2,975 anonymous pairwise votes.
Winner1,472.941
95% confidence interval: ±10.457The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,434.489
95% confidence interval: ±11.462The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlMathematical professions
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 3,546 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 2,828 anonymous pairwise votes.
Winner1,499.27
95% confidence interval: ±10.639The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,445.003
95% confidence interval: ±11.761The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlMedicine and healthcare
Test details
No documented difference in the test setupNo clear leadVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 5,046 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 3,699 anonymous pairwise votes.
No clear lead1,466.482
95% confidence interval: ±9.337The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,459.257
95% confidence interval: ±10.503The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlMedicine and healthcare
Test details
No documented difference in the test setupNo clear leadVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 4,826 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 3,774 anonymous pairwise votes.
No clear lead1,474.929
95% confidence interval: ±9.555The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,461.307
95% confidence interval: ±10.453The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlMulti-turn
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 12,664 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 9,363 anonymous pairwise votes.
Winner1,480.155
95% confidence interval: ±6.57The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,453.414
95% confidence interval: ±7.284The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlMulti-turn
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 11,982 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 9,233 anonymous pairwise votes.
Winner1,492.905
95% confidence interval: ±6.641The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,446.332
95% confidence interval: ±7.34The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlNon-English
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 37,043 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 28,862 anonymous pairwise votes.
Winner1,456.722
95% confidence interval: ±4.588The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,423.239
95% confidence interval: ±4.947The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlNon-English
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 35,486 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 28,837 anonymous pairwise votes.
Winner1,468.109
95% confidence interval: ±4.642The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,423.125
95% confidence interval: ±5.023The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlPolish
Test details
No documented difference in the test setupNo clear leadVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 1,304 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 1,030 anonymous pairwise votes.
No clear lead1,471.709
95% confidence interval: ±16.827The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,438.601
95% confidence interval: ±18.712The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlPolish
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 1,322 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 1,037 anonymous pairwise votes.
Winner1,481.38
95% confidence interval: ±16.86The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,434.977
95% confidence interval: ±18.457The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlRussian
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 7,658 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 5,750 anonymous pairwise votes.
Winner1,469.597
95% confidence interval: ±7.678The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,433.519
95% confidence interval: ±8.502The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlRussian
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 7,218 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 5,790 anonymous pairwise votes.
Winner1,487.106
95% confidence interval: ±7.885The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,437.327
95% confidence interval: ±8.593The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlSoftware and IT services
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 26,958 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 21,360 anonymous pairwise votes.
Winner1,498.74
95% confidence interval: ±5.151The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,470.864
95% confidence interval: ±5.508The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlSoftware and IT services
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 25,454 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 21,329 anonymous pairwise votes.
Winner1,506.553
95% confidence interval: ±5.185The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,472.722
95% confidence interval: ±5.539The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlSpanish
Test details
No documented difference in the test setupNo clear leadVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 2,067 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 1,596 anonymous pairwise votes.
No clear lead1,461.197
95% confidence interval: ±14.33The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,433.099
95% confidence interval: ±16.059The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlSpanish
Test details
No documented difference in the test setupNo clear leadVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 2,086 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 1,497 anonymous pairwise votes.
No clear lead1,454.845
95% confidence interval: ±14.273The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,429.802
95% confidence interval: ±16.45The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlWriting, literature and language
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 16,553 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 12,557 anonymous pairwise votes.
Winner1,458.352
95% confidence interval: ±5.94The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,415.672
95% confidence interval: ±6.557The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026
LMArena Text, Style ControlWriting, literature and language
Test details
No documented difference in the test setupVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control
DeepSeek-V4-Flash: Snapshot: deepseek-v4-flash-high-preview, Reasoning: high, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 16,142 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 12,763 anonymous pairwise votes.
Winner1,464.687
95% confidence interval: ±5.995The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.
1,422.277
95% confidence interval: ±6.549The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.		LMArena
Source details
Data as of: 09/25/2026

Key to the marks:better valuebetter, but not conclusivetie

Our model tests

Our tests

Compare how the models respond to the same prompt. Each test shows the latest attempt, with no subsequent fixes to the generated code. Earlier runs remain available in the test details. These results do not contribute to an overall score.

Pelican on a bicycle

A community classic for free-form SVG drawing.

GPT-5.4

GPT-5.4: Pelican on a bicycle
A community classic for free-form SVG drawing.

Reasoning (requested): high

Current attempt
Run date
1 Oct 2026
API provider
openrouter
Requested API model
openai/gpt-5.4
API model reported in response
openai/gpt-5.4
Reasoning setting
high
Pinned endpoint
OpenAI | openai/gpt-5.4-20260305
Temperature
Provider default
Token limit including reasoning
8,192 tokens
Test protocol
community-visual-tests-v1
Technical checks
Passed, not a quality rating
Response time
116.8 s

Visually reviewed. This individual test does not establish overall model quality.

Earlier attempts (1)
  • 7 Sept 2026API request failedReasoning setting: high

DeepSeek-V4-Flash

The response reached the token limit. No complete recording is available.

Reasoning (requested): high

Current attempt
Run date
2 Oct 2026
API provider
openrouter
Requested API model
deepseek/deepseek-v4-flash-0731
API model reported in response
deepseek/deepseek-v4-flash-0731
Reasoning setting
high
Pinned endpoint
StreamLake | deepseek/deepseek-v4-flash-20260731
Temperature
Provider default
Token limit including reasoning
8,192 tokens
Test protocol
community-visual-tests-v1
Technical checks
Failed
Response time
117.6 s

Not yet visually reviewed. A successful recording does not confirm correct physics or full compliance with the prompt.

Earlier attempts (2)
  • 1 Oct 2026API request failedReasoning setting: high
  • 7 Sept 2026API request failedReasoning setting: high
Prompt and test conditions

Generate an SVG of a pelican riding a bicycle

Token limit including reasoning: 8,192 tokens

Task origin (Simon Willison)

How to read this comparison

Winning one benchmark is not an overall verdict

Your workload, budget, and required context length matter most. A coding benchmark says little about visual understanding or agent performance.

The standard price table uses direct developer rates. The provider comparison labels OpenRouter endpoints, regions, and special tariffs separately.

Every result links to its measurement source and retrieval date. Labels distinguish vendor reports, official benchmark leaderboards, and independent evaluations. Disputed or archived results are excluded.

More matchups

Related LLM comparisons

Compare GPT-5.4 and DeepSeek-V4-Flash with other leading models using the same data and benchmark logic.

View the complete LLM comparison