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: 61
distinct benchmarks: 2
latest retrieval: 2026-09-04

First modelSecond model

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
Attribute	GPT-5.4	DeepSeek-V4-Flash
Provider	OpenAI	DeepSeek
Released	Mar 5, 2026	Apr 24, 2026
Availability	Active	Preview
Model type	Proprietary	Open Weights
Parameters	Not published	284B, 13B active
Architecture	Not published	Mixture of Experts
Context window	1,050,000 tokens	1,000,000 tokens
Knowledge cutoff	August 31, 2025	Not published
Technical sources	Model	Model

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 type	GPT-5.4	DeepSeek-V4-Flash
API input	$2.5 ≤272K / $5 >272K	$0.22
API output	$15 ≤272K / $22.5 >272K	$0.66
Cache read	$0.25 ≤272K / $0.5 >272K	$0.01
Cache write (5 min.)	Not listed	Not listed
Cache write (1 hr.)	Not listed	Not listed
Price verified	07/30/2026 Source	09/08/2026 Source

GPT-5.4DeepSeek-V4-Flash

API inputUSD per 1M tokens, base tier

GPT-5.4$2.5

DeepSeek-V4-Flash$0.22

API outputUSD per 1M tokens, base tier

GPT-5.4$15

DeepSeek-V4-Flash$0.66

Cache readUSD per 1M tokens, base tier

GPT-5.4$0.25

DeepSeek-V4-Flash$0.007

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
Model	Provider	Purchase route and tariff	Input per 1M tokens	Output per 1M tokens	Source
GPT-5.4	OpenAI	Direct from developerUp to 272,000 context tokens	$2.5	$15	Source07/30/2026
	OpenAI	Direct from developerAbove 272,000 context tokens	$5	$22.5	Source07/30/2026
	Amazon Bedrock	Via OpenRouteramazon-bedrock/us-east-1	$2.75	$16.5	Source09/08/2026
	Azure	Via OpenRouterazure	$2.5	$15	Source09/08/2026
	Azure	Via OpenRouterazure/eu	$2.75	$16.5	Source09/08/2026
	Azure	Via OpenRouterazure/us	$2.75	$16.5	Source09/08/2026
	OpenAI	Via OpenRouteropenai	$2.5	$15	Source09/08/2026
	OpenAI	Via OpenRouteropenai/fast	$5	$30	Source09/08/2026
	OpenAI	Via OpenRouteropenai/flex	Lowest listed price: $1.25	Lowest listed price: $7.5	Source09/08/2026
DeepSeek-V4-Flash	DeepSeek	Direct from developerStandard API	$0.22	$0.66	Source09/08/2026
	Alibaba	Via OpenRouteralibaba	$0.352	$1.056	Source09/08/2026
	AtlasCloud	Via OpenRouteratlas-cloud/fp4	$0.44	$1.32	Source09/08/2026
	Baidu	Via OpenRouterbaidu/fp8	$0.14	$0.28	Source09/08/2026
	BaseTen	Via OpenRouterbaseten/fp8#1	$0.13	$0.26	Source09/08/2026
	BaseTen	Via OpenRouterbaseten/fp8#2	$0.13	$0.26	Source09/08/2026
	Cloudflare	Via OpenRoutercloudflare	$0.44	$1.32	Source09/08/2026
	CoreWeave	Via OpenRoutercoreweave/fp8	$0.13	$0.28	Source09/08/2026
	DeepInfra	Via OpenRouterdeepinfra/fp8	$0.06	$0.18	Source09/08/2026
	DeepSeek	Via OpenRouterdeepseek	$0.22	$0.66	Source09/08/2026
	DigitalOcean	Via OpenRouterdigitalocean	$0.08	$0.252	Source09/08/2026
	Fireworks	Via OpenRouterfireworks	$0.22	$0.66	Source09/08/2026
	GMICloud	Via OpenRoutergmicloud/fp8	$0.286	$0.858	Source09/08/2026
	Inceptron	Via OpenRouterinceptron/fp4	$0.13	$0.28	Source09/08/2026
	Makora	Via OpenRoutermakora	$0.09	$0.195	Source09/08/2026
	Mancer 2	Via OpenRoutermancer/fp8	$0.165	$0.5	Source09/08/2026
	Morph	Via OpenRoutermorph/bf16	$0.1234	$0.3475	Source09/08/2026
	NextBit	Via OpenRouternextbit/fp8	$0.352	$1.056	Source09/08/2026
	Novita	Via OpenRouternovita/fp8	$0.4092	$1.2276	Source09/08/2026
	OpenInference	Via OpenRouteropen-inference/fp8	Lowest listed price: $0.05	Lowest listed price: $0.16	Source09/08/2026
	Parasail	Via OpenRouterparasail/fp8	$0.14	$0.28	Source09/08/2026
	Phala	Via OpenRouterphala	$0.44	$1.32	Source09/08/2026
	Reka	Via OpenRouterreka/fp4	$0.11	$0.66	Source09/08/2026
	Relace	Via OpenRouterrelace/fp4	$0.065	$0.18	Source09/08/2026
	Sail Research	Via OpenRoutersail-research/fp4	$0.065	$0.18	Source09/08/2026
	SiliconFlow	Via OpenRoutersiliconflow/fp8	$0.22	$0.66	Source09/08/2026
	StreamLake	Via OpenRouterstreamlake/fp8	$0.132	$0.396	Source09/08/2026
	Together	Via OpenRoutertogether	$0.14	$0.28	Source09/08/2026
	Venice	Via OpenRoutervenice	$0.175	$0.35	Source09/08/2026
	Wafer	Via OpenRouterwafer/fast	$0.1	$0.25	Source09/08/2026

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.

GPT-5.4

DeepSeek-V4-Flash

Coding: 100 / 97.7 (1 benchmark family)

Reasoning: 100 / 97.8 (1 benchmark family)

Mathematics: 100 / 96.9 (1 benchmark family)

Knowledge: 100 / 97.9 (1 benchmark family)

Finance: 100 / 97.1 (1 benchmark family)

Legal: 100 / 98 (1 benchmark family)

Cybersecurity: 100 / 91.8 (1 benchmark family)

Health: 100 / 99.3 (1 benchmark family)

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

GPT-5.4DeepSeek-V4-Flash

CyberBenchOverall. Higher is better. No documented setup difference

GPT-5.472.881 %

DeepSeek-V4-Flash66.716 %

LMArena Text, Style ControlOverall. Higher is better. No documented setup difference

GPT-5.41,465.527

DeepSeek-V4-Flash1,435.816

CyberBenchPatch. Higher is better. No documented setup difference

GPT-5.481.356 %

DeepSeek-V4-Flash72.414 %

LMArena Text, Style ControlOverall. Higher is better. No documented setup difference

GPT-5.41,476.488

DeepSeek-V4-Flash1,438.462

CyberBenchPoC. Higher is better. No documented setup difference

GPT-5.464.407 %

DeepSeek-V4-Flash61.017 %

LMArena Text, Style ControlBusiness, management and finance. Higher is better. No documented setup difference

GPT-5.41,474.871

DeepSeek-V4-Flash1,437.854

LMArena Text, Style ControlBusiness, management and finance. Higher is better. No documented setup difference

GPT-5.41,483.779

DeepSeek-V4-Flash1,434.968

LMArena Text, Style ControlChinese. Higher is better. No documented setup difference

GPT-5.41,494.074

DeepSeek-V4-Flash1,472.677

LMArena Text, Style ControlChinese. Higher is better. No documented setup difference

GPT-5.41,504.962

DeepSeek-V4-Flash1,479.62

LMArena Text, Style ControlCoding. Higher is better. No documented setup difference

GPT-5.41,513.577

DeepSeek-V4-Flash1,483.101

LMArena Text, Style ControlCoding. Higher is better. No documented setup difference

GPT-5.41,519.961

DeepSeek-V4-Flash1,479.449

LMArena Text, Style ControlCreative writing. Higher is better. No documented setup difference

GPT-5.41,436.108

DeepSeek-V4-Flash1,408.18

Key to the marks:better valuebetter, but not conclusivetie

Our model tests

Community tests you can inspect

Compare how the models respond to the same prompt. Each test shows the first attempt, with no subsequent fixes to the generated code. These results do not contribute to an overall score.

Pelican on a bicycle

A community classic for free-form SVG drawing.

GPT-5.4

API access was unavailable for this run. The model was not evaluated.

First attempt

Run date: 7 Sept 2026
API provider: openrouter
Requested API model: openai/gpt-5.4
Pinned endpoint: OpenAI | openai/gpt-5.4-20260305
Reasoning setting: high
Temperature: Provider default
Token limit including reasoning: 8,192 tokens
Test protocol: community-visual-tests-v1
Technical checks: Not run

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

DeepSeek-V4-Flash

API access was unavailable for this run. The model was not evaluated.

First attempt

Run date: 7 Sept 2026
API provider: openrouter
Requested API model: deepseek/deepseek-v4-flash-0731
Pinned endpoint: DeepSeek | deepseek/deepseek-v4-flash-20260731
Reasoning setting: high
Temperature: Provider default
Token limit including reasoning: 8,192 tokens
Test protocol: community-visual-tests-v1
Technical checks: Not run

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

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

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

View the complete LLM comparison

Benchmark scores for GPT-5.4 and DeepSeek-V4-Flash
Benchmark and task	GPT-5.4	DeepSeek-V4-Flash	Source
CyberBenchOverall Test details No documented difference in the test setupVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05	Winner72.881 %	66.716 %	Vals AI Source details Data as of: 09/01/2026
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 63,615 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 48,963 anonymous pairwise votes.	Winner1,465.527 95% confidence interval: ±3.807The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,435.816 95% confidence interval: ±4.15The 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: 09/01/2026
CyberBenchPatch 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-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 60,657 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 48,705 anonymous pairwise votes.	Winner1,476.488 95% confidence interval: ±3.881The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,438.462 95% confidence interval: ±4.21The 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: 09/01/2026
CyberBenchPoC 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 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 13,060 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 9,911 anonymous pairwise votes.	Winner1,474.871 95% confidence interval: ±6.534The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,437.854 95% confidence interval: ±7.128The 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: 09/01/2026
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 12,346 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 10,060 anonymous pairwise votes.	Winner1,483.779 95% confidence interval: ±6.654The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,434.968 95% confidence interval: ±7.142The 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: 09/01/2026
LMArena Text, Style ControlChinese 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 3,684 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 2,530 anonymous pairwise votes.	No clear lead1,494.074 95% confidence interval: ±10.702The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,472.677 95% confidence interval: ±12.483The 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: 09/01/2026
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 3,482 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 2,509 anonymous pairwise votes.	Winner1,504.962 95% confidence interval: ±10.923The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,479.62 95% confidence interval: ±12.459The 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: 09/01/2026
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 17,704 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 14,346 anonymous pairwise votes.	Winner1,513.577 95% confidence interval: ±5.956The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,483.101 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.	LMArena Source details Data as of: 09/01/2026
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 16,441 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 14,310 anonymous pairwise votes.	Winner1,519.961 95% confidence interval: ±6.001The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,479.449 95% confidence interval: ±6.409The 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: 09/01/2026
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 10,326 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 7,972 anonymous pairwise votes.	Winner1,436.108 95% confidence interval: ±7.048The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,408.18 95% confidence interval: ±7.747The 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: 09/01/2026
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 10,108 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 8,029 anonymous pairwise votes.	Winner1,443.763 95% confidence interval: ±7.189The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,408.903 95% confidence interval: ±7.781The 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: 09/01/2026
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 29,058 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 21,931 anonymous pairwise votes.	Winner1,468.527 95% confidence interval: ±4.99The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,445.287 95% confidence interval: ±5.383The 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: 09/01/2026
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 27,549 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 21,747 anonymous pairwise votes.	Winner1,477.92 95% confidence interval: ±5.043The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,449.894 95% confidence interval: ±5.421The 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: 09/01/2026
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 13,621 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 10,423 anonymous pairwise votes.	Winner1,433.464 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.	1,405.349 95% confidence interval: ±7.122The 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: 09/01/2026
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 12,946 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 10,655 anonymous pairwise votes.	Winner1,442.467 95% confidence interval: ±6.606The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,406.152 95% confidence interval: ±7.06The 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: 09/01/2026
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 48,871 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 36,983 anonymous pairwise votes.	Winner1,471.798 95% confidence interval: ±4.961The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,431.271 95% confidence interval: ±5.416The 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: 09/01/2026
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 46,411 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 37,039 anonymous pairwise votes.	Winner1,485.953 95% confidence interval: ±5.055The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,435.7 95% confidence interval: ±5.462The 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: 09/01/2026
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 6,107 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 4,904 anonymous pairwise votes.	Winner1,490.311 95% confidence interval: ±8.598The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,458.425 95% confidence interval: ±9.147The 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: 09/01/2026
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 5,765 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 4,907 anonymous pairwise votes.	Winner1,515.538 95% confidence interval: ±8.748The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,467.744 95% confidence interval: ±9.211The 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: 09/01/2026
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 2,327 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 1,909 anonymous pairwise votes.	Winner1,484.932 95% confidence interval: ±14.655The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,444.246 95% confidence interval: ±15.542The 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: 09/01/2026
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 2,288 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 1,872 anonymous pairwise votes.	Winner1,502.194 95% confidence interval: ±14.8The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,457.512 95% confidence interval: ±15.76The 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: 09/01/2026
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 964 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 834 anonymous pairwise votes.	No clear lead1,448.071 95% confidence interval: ±20.361The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,423.985 95% confidence interval: ±21.432The 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: 09/01/2026
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-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 1,056 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 881 anonymous pairwise votes.	No clear lead1,470.385 95% confidence interval: ±19.432The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,434.001 95% confidence interval: ±20.992The 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: 09/01/2026
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 41,790 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 32,576 anonymous pairwise votes.	Winner1,487.567 95% confidence interval: ±4.591The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,459.19 95% confidence interval: ±4.879The 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: 09/01/2026
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 39,303 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 32,479 anonymous pairwise votes.	Winner1,497.941 95% confidence interval: ±4.628The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,458.492 95% confidence interval: ±4.934The 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: 09/01/2026
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 19,850 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 15,327 anonymous pairwise votes.	Winner1,487.46 95% confidence interval: ±5.732The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,464.306 95% confidence interval: ±6.191The 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: 09/01/2026
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 18,552 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 15,135 anonymous pairwise votes.	Winner1,495.34 95% confidence interval: ±5.797The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,465.194 95% confidence interval: ±6.261The 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: 09/01/2026
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 21,742 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 16,772 anonymous pairwise votes.	Winner1,461.072 95% confidence interval: ±5.529The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,428.139 95% confidence interval: ±5.955The 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: 09/01/2026
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 20,284 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 16,813 anonymous pairwise votes.	Winner1,472.675 95% confidence interval: ±5.598The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,433.908 95% confidence interval: ±6.024The 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: 09/01/2026
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 652 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 633 anonymous pairwise votes.	No clear lead1,445.14 95% confidence interval: ±24.878The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,397.389 95% confidence interval: ±25.351The 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: 09/01/2026
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 644 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 600 anonymous pairwise votes.	Winner1,482.947 95% confidence interval: ±25.629The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,422.217 95% confidence interval: ±25.99The 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: 09/01/2026
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 1,132 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 867 anonymous pairwise votes.	No clear lead1,426.207 95% confidence interval: ±19.624The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,394.225 95% confidence interval: ±21.434The 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: 09/01/2026
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 1,077 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 909 anonymous pairwise votes.	Winner1,437.428 95% confidence interval: ±20.212The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,374.943 95% confidence interval: ±21.551The 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: 09/01/2026
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 5,056 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 3,892 anonymous pairwise votes.	Winner1,476.119 95% confidence interval: ±9.309The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,449.174 95% confidence interval: ±10.185The 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: 09/01/2026
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 4,896 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 3,872 anonymous pairwise votes.	Winner1,490.489 95% confidence interval: ±9.43The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,459.091 95% confidence interval: ±10.242The 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: 09/01/2026
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 10,310 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 7,776 anonymous pairwise votes.	Winner1,478.714 95% confidence interval: ±6.907The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,458.608 95% confidence interval: ±7.61The 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: 09/01/2026
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 9,887 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 7,968 anonymous pairwise votes.	Winner1,487.819 95% confidence interval: ±7.029The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,460.882 95% confidence interval: ±7.588The 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: 09/01/2026
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 27,399 anonymous pairwise votes.DeepSeek-V4-Flash: Style-controlled Arena rating from 21,561 anonymous pairwise votes.	Winner1,478.044 95% confidence interval: ±5.365The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result.	1,448.456 95% confidence interval: ±5.733The 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: 09/01/2026