Head-to-head comparison

GPT-5.4 vs. Gemini 3.8 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: 186
distinct benchmarks: 19
latest retrieval: 2026-10-15

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 Gemini 3.8 Flash
Attribute	GPT-5.4	Gemini 3.8 Flash
Provider	OpenAI	Google
Released	Mar 5, 2026	Sep 2, 2026
Availability	Active	Active
Model type	Proprietary	Proprietary
Parameters	Not published	Not published
Architecture	Not published	Not published
Context window	1,050,000 tokens	1,048,576 tokens
Knowledge cutoff	August 31, 2025	March 2026
Technical sources	Model	Model, Knowledge cutoff

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 Gemini 3.8 Flash
Price type	GPT-5.4	Gemini 3.8 Flash
API input	$2.5 ≤272K / $5 >272K	$0.75
API output	$15 ≤272K / $22.5 >272K	$3.75
Cache read	$0.25 ≤272K / $0.5 >272K	$0.075
Cache write (5 min.)	Not listed	Not listed
Cache write (1 hr.)	Not listed	Not listed
Price verified	09/22/2026 Source	09/22/2026 Source

GPT-5.4Gemini 3.8 Flash

API inputUSD per 1M tokens, base tier

GPT-5.4$2.5

Gemini 3.8 Flash$0.75

API outputUSD per 1M tokens, base tier

GPT-5.4$15

Gemini 3.8 Flash$3.75

Cache readUSD per 1M tokens, base tier

GPT-5.4$0.25

Gemini 3.8 Flash$0.075

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 Gemini 3.8 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	Source09/22/2026
	OpenAI	Direct from developerAbove 272,000 context tokens	$5	$22.5	Source09/22/2026
	Amazon Bedrock	Via OpenRouteramazon-bedrock/us-east-1	$2.75	$16.5	Source10/18/2026
	Azure	Via OpenRouterazure	$2.5	$15	Source10/18/2026
	Azure	Via OpenRouterazure/eu	$2.75	$16.5	Source10/18/2026
	Azure	Via OpenRouterazure/us	$2.75	$16.5	Source10/18/2026
	OpenAI	Via OpenRouteropenai	$2.5	$15	Source10/18/2026
	OpenAI	Via OpenRouteropenai/fast	$5	$30	Source10/18/2026
	OpenAI	Via OpenRouteropenai/flex	Lowest listed price: $1.25	Lowest listed price: $7.5	Source10/18/2026
Gemini 3.8 Flash	Google	Direct from developerStandard API	$0.75	$3.75	Source09/22/2026
	Google	Via OpenRoutergoogle-vertex/global	$0.75	$3.75	Source10/18/2026
	Google	Via OpenRoutergoogle-vertex/global/flex	Lowest listed price: $0.375	Lowest listed price: $1.875	Source10/18/2026
	Google	Via OpenRoutergoogle-vertex/global/priority	$1.35	$6.75	Source10/18/2026
	Google	Via OpenRoutergoogle-vertex/us	$0.825	$4.125	Source10/18/2026
	Google AI Studio	Via OpenRoutergoogle-ai-studio	$0.75	$3.75	Source10/18/2026
	Google AI Studio	Via OpenRoutergoogle-ai-studio/flex	Lowest listed price: $0.375	Lowest listed price: $1.875	Source10/18/2026
	Google AI Studio	Via OpenRoutergoogle-ai-studio/priority	$1.35	$6.75	Source10/18/2026

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

Gemini 3.8 Flash

Coding: 99.5 / 100 (1 benchmark family)

Reasoning: 98.9 / 100 (1 benchmark family)

Mathematics: 98.6 / 100 (1 benchmark family)

Knowledge: 98.1 / 100 (1 benchmark family)

Finance: 100 / 99.9 (1 benchmark family)

Legal: 99.9 / 100 (1 benchmark family)

Multimodal: 98.8 / 100 (1 benchmark family)

Health: 97.7 / 100 (1 benchmark family)

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

GPT-5.4Gemini 3.8 Flash

Code MigrationOverall. Higher is better. Test setup not documented as identical

GPT-5.434.984 %

Gemini 3.8 Flash36.546 %

GPQA DiamondOverall. Higher is better. Test setup not documented as identical

GPT-5.491.666 %

Gemini 3.8 Flash94.444 %

Harvey's Legal Agent BenchmarkOverall · Task fully resolved. Higher is better. Test setup not documented as identical

GPT-5.40 %

Gemini 3.8 Flash10 %

LegalBenchOverall. Higher is better. Test setup not documented as identical

GPT-5.486.044 %

Gemini 3.8 Flash86.993 %

LiveCodeBench v6Overall. Higher is better. Test setup not documented as identical

GPT-5.484.141 %

Gemini 3.8 Flash89.484 %

LMArena Code Arena WebDevOverall. Higher is better. Test setup not documented as identical

GPT-5.41,401.766

Gemini 3.8 Flash1,584.77

LMArena Text, Style ControlOverall. Higher is better. Test setup not documented as identical

GPT-5.41,464.728

Gemini 3.8 Flash1,496.319

LMArena Vision, Style ControlOverall. Higher is better. Test setup not documented as identical

GPT-5.41,278.895

Gemini 3.8 Flash1,289.129

MedCodeOverall. Higher is better. Test setup not documented as identical

GPT-5.441.292 %

Gemini 3.8 Flash48.135 %

MedScribeOverall. Higher is better. Test setup not documented as identical

GPT-5.477.549 %

Gemini 3.8 Flash84.496 %

MMLU-ProOverall. Higher is better. Test setup not documented as identical

GPT-5.487.482 %

Gemini 3.8 Flash90.221 %

MMMU-ProOverall. Higher is better. Test setup not documented as identical

GPT-5.487.514 %

Gemini 3.8 Flash89.075 %

Key to the marks:better valuebetter, but not conclusivetie

Benchmark scores for GPT-5.4 and Gemini 3.8 Flash
Benchmark and task	GPT-5.4	Gemini 3.8 Flash	Source
Code MigrationOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	34.984 %	Higher measured value36.546 %	Vals AI Source details Data as of: 10/12/2026
GPQA DiamondOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	91.666 %	Higher measured value94.444 %	Vals AI Source details Data as of: 09/01/2026
Harvey's Legal Agent BenchmarkOverall · Task fully resolved Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: task resolution rateScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	0 %	Higher measured value10 %	Vals AI Source details Data as of: 10/12/2026
LegalBenchOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	86.044 %	Higher measured value86.993 %	Vals AI Source details Data as of: 10/13/2026
LiveCodeBench v6Overall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	84.141 %	Higher measured value89.484 %	Vals AI Source details Data as of: 09/01/2026
LMArena Code Arena WebDevOverall Test details Test setup not documented as identicalNo direct winnerVersion: webdevMetric: WebDev Arena Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Code Arena WebDev Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Code Arena WebDevGPT-5.4: Arena rating from 3,990 published pairwise votes.Gemini 3.8 Flash: Arena rating from 10,077 published pairwise votes.	1,401.766 95% confidence interval: ±10.455The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,584.77 95% confidence interval: ±7.59The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/13/2026
LMArena Text, Style ControlOverall Test details Test setup not documented as identicalNo direct winnerVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Text, style control Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Text, style controlGPT-5.4: Arena rating from 67,880 published pairwise votes.Gemini 3.8 Flash: Arena rating from 35,670 published pairwise votes.	1,464.728 95% confidence interval: ±3.69The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,496.319 95% confidence interval: ±4.654The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/13/2026
LMArena Vision, Style ControlOverall Test details Test setup not documented as identicalNo direct winnerVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Vision, style control Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Vision, style controlGPT-5.4: Arena rating from 22,832 published pairwise votes.Gemini 3.8 Flash: Arena rating from 3,792 published pairwise votes.	1,278.895 95% confidence interval: ±6.536The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,289.129 95% confidence interval: ±10.62The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/09/2026
MedCodeOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 30,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	41.292 %	Higher measured value48.135 %	Vals AI Source details Data as of: 10/13/2026
MedScribeOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 30,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	77.549 %	Higher measured value84.496 %	Vals AI Source details Data as of: 10/13/2026
MMLU-ProOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	87.482 %	Higher measured value90.221 %	Vals AI Source details Data as of: 09/01/2026
MMMU-ProOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	87.514 %	Higher measured value89.075 %	Vals AI Source details Data as of: 09/01/2026
MortgageTaxOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	Higher measured value68.323 %	65.342 %	Vals AI Source details Data as of: 09/01/2026
ProgramBenchOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: fully resolved rateScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	0 %	Higher measured value1 %	Vals AI Source details Data as of: 10/12/2026
SAGEOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 30,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 30,000 tokens	Higher measured value43.312 %	35.064 %	Vals AI Source details Data as of: 10/08/2026
SkillsBenchOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Harness: OpenHands, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Harness: OpenHands, Temperature: 1, Output limit: 65,536 tokens	51.711 %	Higher measured value57.985 %	Vals AI Source details Data as of: 09/27/2026
SWE-bench VerifiedOverall Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	78.2 %	Higher measured value80 %	Vals AI Source details Data as of: 09/01/2026
TaxEval v2Overall Test details Test setup not documented as identicalNo direct winnerVersion: 2Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	73.958 %	Higher measured value74.448 %	Vals AI Source details Data as of: 09/01/2026
Vibe Code Bench v1.1Overall Test details Test setup not documented as identicalNo direct winnerVersion: 1.1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Harness: OpenHands, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Harness: OpenHands, Temperature: 1, Output limit: 65,536 tokens	67.421 %	Higher measured value78.651 %	Vals AI Source details Data as of: 10/12/2026
Code MigrationCLI Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	23.788 %	Higher measured value25.871 %	Vals AI Source details Data as of: 10/12/2026
GPQA DiamondFew-Shot CoT Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	91.414 %	Higher measured value94.949 %	Vals AI Source details Data as of: 09/01/2026
Harvey's Legal Agent BenchmarkAntitrust Competition · Criteria passed Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: criteria pass rateScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	76.449 %	Higher measured value87.681 %	Vals AI Source details Data as of: 10/12/2026
LegalBenchConclusion Tasks Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	88.939 %	Higher measured value90.691 %	Vals AI Source details Data as of: 10/13/2026
LiveCodeBench v6Easy Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	No direct winner99.068 %	No direct winner99.068 %	Vals AI Source details Data as of: 09/01/2026
LMArena Code Arena WebDevImage to webdev Test details Test setup not documented as identicalNo direct winnerVersion: webdevMetric: WebDev Arena Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Code Arena WebDev Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Code Arena WebDevGPT-5.4: Arena rating from 3,076 published pairwise votes.Gemini 3.8 Flash: Arena rating from 2,015 published pairwise votes.	1,440.409 95% confidence interval: ±11.975The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,593.067 95% confidence interval: ±13.862The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/09/2026
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 Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Text, style controlGPT-5.4: Arena rating from 64,787 published pairwise votes.Gemini 3.8 Flash: Arena rating from 35,670 published pairwise votes.	1,475.066 95% confidence interval: ±3.765The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Winner1,496.319 95% confidence interval: ±4.654The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/13/2026
LMArena Vision, Style ControlOverall 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 Vision, style control Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Vision, style controlGPT-5.4: Arena rating from 29,881 published pairwise votes.Gemini 3.8 Flash: Arena rating from 3,792 published pairwise votes.	1,284.987 95% confidence interval: ±5.834The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	No clear lead1,289.129 95% confidence interval: ±10.62The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/09/2026
MMLU-ProBiology Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	92.887 %	Higher measured value94.979 %	Vals AI Source details Data as of: 09/01/2026
MortgageTaxNumerical Extraction Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	Higher measured value75.994 %	70.588 %	Vals AI Source details Data as of: 09/01/2026
ProgramBenchAlmost Resolved Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: almost-resolved rateScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	7 %	Higher measured value14 %	Vals AI Source details Data as of: 10/12/2026
SAGEAutomata Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 30,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 30,000 tokens	21.428 %	Higher measured value33.334 %	Vals AI Source details Data as of: 10/08/2026
SkillsBenchNo Skills Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Harness: OpenHands, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Harness: OpenHands, Temperature: 1, Output limit: 65,536 tokens	Higher measured value48.905 %	41.046 %	Vals AI Source details Data as of: 09/27/2026
SWE-bench Verified<15 min fix Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	88.144 %	Higher measured value89.691 %	Vals AI Source details Data as of: 09/01/2026
TaxEval v2Correctness Test details Test setup not documented as identicalNo direct winnerVersion: 2Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	Higher measured value64.922 %	62.469 %	Vals AI Source details Data as of: 09/01/2026
Code MigrationCOBOL Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	No direct winner68.571 %	No direct winner68.571 %	Vals AI Source details Data as of: 10/12/2026
GPQA DiamondZero-Shot CoT Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	91.919 %	Higher measured value93.939 %	Vals AI Source details Data as of: 09/01/2026
Harvey's Legal Agent BenchmarkAntitrust Competition · Task fully resolved Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: task resolution rateScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	No direct winner0 %	No direct winner0 %	Vals AI Source details Data as of: 10/12/2026
LegalBenchInterpretation Tasks Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	83.921 %	Higher measured value84.589 %	Vals AI Source details Data as of: 10/13/2026
LiveCodeBench v6Hard Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	67.714 %	Higher measured value78 %	Vals AI Source details Data as of: 09/01/2026
LMArena Code Arena WebDevWebdev Test details Test setup not documented as identicalNo direct winnerVersion: webdevMetric: WebDev Arena Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Code Arena WebDev Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Code Arena WebDevGPT-5.4: Arena rating from 3,990 published pairwise votes.Gemini 3.8 Flash: Arena rating from 10,077 published pairwise votes.	1,401.766 95% confidence interval: ±10.455The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,584.77 95% confidence interval: ±7.59The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/13/2026
LMArena Text, Style ControlBusiness, management and finance Test details Test setup not documented as identicalNo direct winnerVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Text, style control Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Text, style controlGPT-5.4: Arena rating from 13,848 published pairwise votes.Gemini 3.8 Flash: Arena rating from 6,520 published pairwise votes.	1,473.882 95% confidence interval: ±6.361The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,479.887 95% confidence interval: ±8.136The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/13/2026
LMArena Vision, Style ControlChinese Test details Test setup not documented as identicalNo direct winnerVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Vision, style control Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Vision, style controlGPT-5.4: Arena rating from 1,357 published pairwise votes.Gemini 3.8 Flash: Arena rating from 182 published pairwise votes.	1,297.645 95% confidence interval: ±18.927The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,379.729 95% confidence interval: ±43.521The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/09/2026
MMLU-ProBusiness Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	91.762 %	Higher measured value93.79 %	Vals AI Source details Data as of: 09/01/2026
MortgageTaxSemantic Extraction Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	Higher measured value60.652 %	60.095 %	Vals AI Source details Data as of: 09/01/2026
ProgramBenchFully Resolved Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: fully resolved rateScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	0 %	Higher measured value1 %	Vals AI Source details Data as of: 10/12/2026
SAGECalculus Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 30,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 30,000 tokens	70.988 %	Higher measured value79.63 %	Vals AI Source details Data as of: 10/08/2026
SkillsBenchWith Skills Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Harness: OpenHands, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Harness: OpenHands, Temperature: 1, Output limit: 65,536 tokens	51.711 %	Higher measured value57.985 %	Vals AI Source details Data as of: 09/27/2026
SWE-bench Verified>4 hours Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	0 %	Higher measured value33.333 %	Vals AI Source details Data as of: 09/01/2026
TaxEval v2Stepwise Reasoning Test details Test setup not documented as identicalNo direct winnerVersion: 2Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	82.993 %	Higher measured value86.427 %	Vals AI Source details Data as of: 09/01/2026
Code MigrationCode Quality Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: mean reviewer code-quality ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	3.632 / 5	Higher measured value3.742 / 5	Vals AI Source details Data as of: 10/12/2026
Harvey's Legal Agent BenchmarkArbitration International Dispute Resolution · Criteria passed Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: criteria pass rateScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	66.667 %	Higher measured value89.606 %	Vals AI Source details Data as of: 10/12/2026
LegalBenchIssue Tasks Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	90.526 %	Higher measured value91.278 %	Vals AI Source details Data as of: 10/13/2026
LiveCodeBench v6Medium Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	85.64 %	Higher measured value91.384 %	Vals AI Source details Data as of: 09/01/2026
LMArena Code Arena WebDevWebdev-html Test details Test setup not documented as identicalNo direct winnerVersion: webdevMetric: WebDev Arena Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Code Arena WebDev Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Code Arena WebDevGPT-5.4: Arena rating from 528 published pairwise votes.Gemini 3.8 Flash: Arena rating from 1,451 published pairwise votes.	1,434.816 95% confidence interval: ±27.294The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,620.104 95% confidence interval: ±17.559The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/13/2026
LMArena Text, Style ControlBusiness, management and finance 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 Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Text, style controlGPT-5.4: Arena rating from 13,095 published pairwise votes.Gemini 3.8 Flash: Arena rating from 6,520 published pairwise votes.	No clear lead1,481.483 95% confidence interval: ±6.462The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	1,479.887 95% confidence interval: ±8.136The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/13/2026
LMArena Vision, 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 Vision, style control Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Vision, style controlGPT-5.4: Arena rating from 1,807 published pairwise votes.Gemini 3.8 Flash: Arena rating from 182 published pairwise votes.	1,316.134 95% confidence interval: ±16.726The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Winner1,379.729 95% confidence interval: ±43.521The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/09/2026
MMLU-ProChemistry Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	91.078 %	Higher measured value92.491 %	Vals AI Source details Data as of: 09/01/2026
ProgramBenchRaw Pass Rate Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: raw test pass rateScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	68.596 %	Higher measured value71.899 %	Vals AI Source details Data as of: 10/12/2026
SAGELinear Algebra Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 30,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 30,000 tokens	Higher measured value68.826 %	41.7 %	Vals AI Source details Data as of: 10/08/2026
SWE-bench Verified1-4 hours Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	50 %	Higher measured value52.381 %	Vals AI Source details Data as of: 09/01/2026
Harvey's Legal Agent BenchmarkArbitration International Dispute Resolution · Task fully resolved Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: task resolution rateScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	No direct winner0 %	No direct winner0 %	Vals AI Source details Data as of: 10/12/2026
LegalBenchRhetoric Tasks Test details Test setup not documented as identicalNo direct winnerVersion: 1Metric: accuracyScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.8 Flash: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens	Higher measured value82.824 %	81.333 %	Vals AI Source details Data as of: 10/13/2026
LMArena Code Arena WebDevWebdev-react Test details Test setup not documented as identicalNo direct winnerVersion: webdevMetric: WebDev Arena Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Code Arena WebDev Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Code Arena WebDevGPT-5.4: Arena rating from 2,727 published pairwise votes.Gemini 3.8 Flash: Arena rating from 6,460 published pairwise votes.	1,391.351 95% confidence interval: ±12.752The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,576.599 95% confidence interval: ±8.938The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/13/2026
LMArena Text, Style ControlChinese Test details Test setup not documented as identicalNo direct winnerVersion: style-control-v1Metric: style-controlled Bradley-Terry ratingScoring: Higher is betterGPT-5.4: Snapshot: gpt-5.4, Harness: Arena Text, style control Gemini 3.8 Flash: Snapshot: gemini-3.8-flash-high, Reasoning: high, Harness: Arena Text, style controlGPT-5.4: Arena rating from 4,351 published pairwise votes.Gemini 3.8 Flash: Arena rating from 2,903 published pairwise votes.	1,493.612 95% confidence interval: ±9.825The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	Higher measured value1,532.833 95% confidence interval: ±11.757The Arena score measures human preference rather than a fixed capability test. The confidence interval, Arena category, and model configuration are part of the result.	LMArena Source details Data as of: 10/13/2026
Key to the marks:better valuebetter, but not conclusivetie

Our model tests

Our tests

SVG short test

One SVG prompt and fixed output settings. Captures are shown only with 0 reported reasoning tokens or documented lack of reasoning controls. Identical test conditions do not make the models' internal compute identical.

Shared test round: 4ae5d8eeb9a9

GPT-5.4

Unavailable in this round

Gemini 3.8 Flash

Unavailable in this round

Earlier tests

Pelican on a bicycle

A community classic for free-form SVG drawing.

GPT-5.4

Reasoning (requested): high

Current attempt

Run date: Oct 1, 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)

Sep 7, 2026API request failedReasoning setting: high

Gemini 3.8 Flash

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

Reasoning (requested): high

Current attempt

Run date: Oct 5, 2026
API provider: openrouter
Requested API model: google/gemini-3.8-flash
API model reported in response: google/gemini-3.8-flash
Reasoning setting: high
Pinned endpoint: Google AI Studio | google/gemini-3.8-flash-20260902
Temperature: Provider default
Token limit including reasoning: 8,192 tokens
Test protocol: community-visual-tests-v1
Technical checks: Failed
Response time: 40.2 s

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

Earlier attempts (3)

Oct 2, 2026API request failedReasoning setting: high
Oct 1, 2026API request failedReasoning setting: high
Sep 7, 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

Compare GPT-5.4 and Gemini 3.8 Flash with other leading models using the same data and benchmark logic.

View the complete LLM comparison

GPT-5.4 vs. Gemini 3.8 Flash

Change model selection

General specifications

API pricing compared

Shared benchmarks

Capability profile from matched benchmarks

Our tests

SVG short test

GPT-5.4

Gemini 3.8 Flash

Pelican on a bicycle

Winning one benchmark is not an overall verdict

Related LLM comparisons