Head-to-head comparison
GPT-5.4 vs. Gemini 3.1 Pro Preview
Compare published benchmark results from matching versions and cohorts, API pricing, and technical specifications. Any documented test setup differences remain visible.
138
shared result series
25
distinct benchmarks
2026-08-16
latest retrieval
Change model selection
Specifications and cost
General specifications and standard API pricing
Undisclosed parameter counts and knowledge cutoffs stay marked as unavailable. The table shows direct standard token pricing, including documented context tiers. Batch, fast, priority, regional surcharges, tool calls, and cloud platform pricing are excluded.
| Attribute | GPT-5.4 | Gemini 3.1 Pro Preview |
|---|---|---|
| Provider | OpenAI | |
| Released | Mar 2026 | Feb 19, 2026 |
| Availability | Active | Preview |
| 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 | Not published |
| Technical sources | Model | Model |
| API pricing per 1M tokens | ||
| API input | $2.5 ≤272K / $5 >272K | $2 ≤200K / $4 >200K |
| API output | $15 ≤272K / $22.5 >272K | $12 ≤200K / $18 >200K |
| Cache read | $0.25 ≤272K / $0.5 >272K | $0.2 ≤200K / $0.4 >200K |
| Cache write (5 min.) | Not listed | Not listed |
| Cache write (1 hr.) | Not listed | Not listed |
| Price verified | 07/30/2026Source | 07/30/2026Source |
GPT-5.4Gemini 3.1 Pro Preview
API inputUSD per 1M tokens, base tier
GPT-5.4$2.5
Gemini 3.1 Pro Preview$2Lower price
API outputUSD per 1M tokens, base tier
GPT-5.4$15
Gemini 3.1 Pro Preview$12Lower price
Cache readUSD per 1M tokens, base tier
GPT-5.4$0.25
Gemini 3.1 Pro Preview$0.2Lower price
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.
138 shared result series
GPT-5.4Gemini 3.1 Pro Preview
Arena Text, Style ControlOverall. Higher is better. No documented setup difference
GPT-5.41,465.439 arena points
Gemini 3.1 Pro Preview1,486.245 arena pointsWinner
Code MigrationOverall. Higher is better. No documented setup difference
GPT-5.434.984 %Winner
Gemini 3.1 Pro Preview17.313 %
CorpFin v2Overall. Higher is better. No documented setup difference
GPT-5.465.268 %Winner
Gemini 3.1 Pro Preview64.491 %
CyberBenchOverall. Higher is better. No documented setup difference
GPT-5.472.881 %Winner
Gemini 3.1 Pro Preview36.441 %
GPQA DiamondOverall. Higher is better. No documented setup difference
GPT-5.491.666 %
Gemini 3.1 Pro Preview95.454 %Winner
Harvey's Legal Agent BenchmarkOverall · Task fully resolved. Higher is better. No documented setup difference
GPT-5.40 %Tie
Gemini 3.1 Pro Preview0 %Tie
Humanity's Last ExamOverall. Higher is better. No documented setup difference
GPT-5.436.24% ± 1.88
Gemini 3.1 Pro Preview46.44% ± 1.96Winner
LegalBenchOverall. Higher is better. No documented setup difference
GPT-5.486.044 %
Gemini 3.1 Pro Preview87.398 %Winner
LiveCodeBench v6Overall. Higher is better. No documented setup difference
GPT-5.484.141 %
Gemini 3.1 Pro Preview88.485 %Winner
MedCodeOverall. Higher is better. No documented setup difference
GPT-5.441.292 %
Gemini 3.1 Pro Preview59.062 %Winner
MedQAOverall. Higher is better. No documented setup difference
GPT-5.496.092 %
Gemini 3.1 Pro Preview96.367 %Winner
MedScribeOverall. Higher is better. No documented setup difference
GPT-5.477.549 %Winner
Gemini 3.1 Pro Preview76.114 %
| Benchmark and task | GPT-5.4 | Gemini 3.1 Pro Preview | Comparability | Source |
|---|---|---|---|---|
| Arena Text, Style ControlOverall | 1,465.439 arena points 95% confidence interval: ±3.814The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result. | 1,486.245 arena pointsWinner 95% confidence interval: ±3.337The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result. | style-control-v1style-controlled Bradley-Terry ratingHigher is betterCohort: lmarena-text-style-control:2026-08-12:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control Gemini 3.1 Pro Preview: Snapshot: gemini-3.1-pro-preview, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 63,759 anonymous pairwise votes.Gemini 3.1 Pro Preview: Style-controlled Arena rating from 94,989 anonymous pairwise votes. | ArenaType: Official benchmark leaderboardLeaderboard updated: 2026-08-12Import ID: lmarena-eeedf9209373a7f9b324bfb839e7b02fedacd1e5-1611ed64f3f1Retrieved 2026-08-16 |
| Code MigrationOverall | 34.984 %Winner | 17.313 % | 1accuracyHigher is betterCohort: vals-code-migration:1:overallNo documented setup differenceWinner: GPT-5.4GPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-13Import ID: vals-code-migration-2026-08-16-107652dc6bfcRetrieved 2026-08-16 |
| CorpFin v2Overall | 65.268 %Winner | 64.491 % | 2accuracyHigher is betterCohort: vals-corp-fin-v2:2:overallNo documented setup differenceWinner: GPT-5.4GPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-corp_fin_v2-2026-08-16-45bbf84f57abRetrieved 2026-08-16 |
| CyberBenchOverall | 72.881 %Winner | 36.441 % | 1accuracyHigher is betterCohort: vals-cyberbench:1:overallNo documented setup differenceWinner: GPT-5.4GPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-03Import ID: vals-cyber-2026-08-16-702acaee9aa9Retrieved 2026-08-16 |
| GPQA DiamondOverall | 91.666 % | 95.454 %Winner | 1accuracyHigher is betterCohort: vals-gpqa:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-gpqa-2026-08-16-7f8363eaa2dbRetrieved 2026-08-16 |
| Harvey's Legal Agent BenchmarkOverall · Task fully resolved | 0 %Tie | 0 %Tie | 1task resolution rateHigher is betterCohort: vals-legal-agent-benchmark:1:overallNo documented setup differenceTieGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-14Import ID: vals-hlab-2026-08-16-bae90566df2fRetrieved 2026-08-16 |
| Humanity's Last ExamOverall | 36.24% ± 1.88 95% confidence interval: ±1.88 | 46.44% ± 1.96Winner 95% confidence interval: ±1.96 | final 2,500-question public setaccuracyHigher is betterCohort: hle-final-scale-public-temperature-0No documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Attempts: pass@1 Gemini 3.1 Pro Preview: Snapshot: gemini-3.1-pro-preview-thinking-high, Attempts: pass@1All 2,500 public questions, temperature 0, and an automatic answer extractor. | Scale AIType: Official benchmark leaderboardImport ID: curated-scale-ai-2026-08-02-9f109d68fc81 / curated-scale-ai-2026-08-02-606cf1619b66Retrieved 2026-08-02 |
| LegalBenchOverall | 86.044 % | 87.398 %Winner | 1accuracyHigher is betterCohort: vals-legal-bench:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-legal_bench-2026-08-16-d5b3b1b2b5c8Retrieved 2026-08-16 |
| LiveCodeBench v6Overall | 84.141 % | 88.485 %Winner | 1accuracyHigher is betterCohort: vals-lcb:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-13Import ID: vals-lcb-2026-08-16-17b9c147080dRetrieved 2026-08-16 |
| MedCodeOverall | 41.292 % | 59.062 %Winner | 1accuracyHigher is betterCohort: vals-medcode:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-medcode-2026-08-16-524f7f474eacRetrieved 2026-08-16 |
| MedQAOverall | 96.092 % | 96.367 %Winner | 1accuracyHigher is betterCohort: vals-medqa:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-04-16Import ID: vals-medqa-2026-08-16-4d228d35a54bRetrieved 2026-08-16 |
| MedScribeOverall | 77.549 %Winner | 76.114 % | 1accuracyHigher is betterCohort: vals-medscribe:1:overallNo documented setup differenceWinner: GPT-5.4GPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-medscribe-2026-08-16-c385354a7e62Retrieved 2026-08-16 |
| MMLU-ProOverall | 87.482 % | 90.987 %Winner | 1accuracyHigher is betterCohort: vals-mmlu-pro:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-mmlu_pro-2026-08-16-e14d5bb3a34cRetrieved 2026-08-16 |
| MMMU-ProOverall | 87.514 % | 88.208 %Winner | 1accuracyHigher is betterCohort: vals-mmmu:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-mmmu-2026-08-16-37c9a8c7e21fRetrieved 2026-08-16 |
| MortgageTaxOverall | 68.323 % | 69.396 %Winner | 1accuracyHigher is betterCohort: vals-mortgage-tax:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-mortgage_tax-2026-08-16-e78c985dbb1cRetrieved 2026-08-16 |
| MultiNRCOverall | 58.29% ± 2.98 95% confidence interval: ±2.98 | 64.74% ± 2.88Higher measured value 95% confidence interval: ±2.88 | French, Spanish, and Chinese releaseaccuracyHigher is betterCohort: scale-multinrc-native-languagesTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-xhigh, Reasoning: xhigh, Attempts: pass@1 Gemini 3.1 Pro Preview: Attempts: pass@1Native, non-translated reasoning tasks with automatic evaluation. | Scale AIType: Official benchmark leaderboardImport ID: curated-scale-ai-2026-08-05-5ba6c48ecc82 / curated-scale-ai-2026-08-05-5a1557803da7Retrieved 2026-08-05 |
| PRBench FinanceOverall | 45.63% ± 0.56Higher measured value 95% confidence interval: ±0.56 | 41.87% ± 1.22 95% confidence interval: ±1.22 | full finance setweighted rubric scoreHigher is betterCohort: scale-prbench-finance-fullTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Attempts: pass@1 Gemini 3.1 Pro Preview: Attempts: pass@1600 finance tasks with task-specific rubrics and o4-mini as judge. | Scale AIType: Official benchmark leaderboardImport ID: curated-scale-ai-2026-08-05-f6b08a569deb / curated-scale-ai-2026-08-05-c33e65ae0f0eRetrieved 2026-08-05 |
| PRBench LegalOverall | 44.35% ± 0.28Higher measured value 95% confidence interval: ±0.28 | 44.02% ± 1.12 95% confidence interval: ±1.12 | full legal setweighted rubric scoreHigher is betterCohort: scale-prbench-legal-fullTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Attempts: pass@1 Gemini 3.1 Pro Preview: Attempts: pass@1500 legal tasks with task-specific rubrics and o4-mini as judge. | Scale AIType: Official benchmark leaderboardImport ID: curated-scale-ai-2026-08-05-63e40da2f52b / curated-scale-ai-2026-08-05-ddea150189a5Retrieved 2026-08-05 |
| ProgramBenchOverall | 0 %Tie | 0 %Tie | 1accuracyHigher is betterCohort: vals-programbench:1:overallNo documented setup differenceTieGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-13Import ID: vals-programbench-2026-08-16-08d9727662ddRetrieved 2026-08-16 |
| SAGEOverall | 43.312 % | 48.677 %Winner | 1accuracyHigher is betterCohort: vals-sage:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-sage-2026-08-16-a416dd25affbRetrieved 2026-08-16 |
| SWE-bench VerifiedOverall | 78.2 % | 78.8 %Winner | 1accuracyHigher is betterCohort: vals-swebench:1:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-swebench-2026-08-16-c133071b8167Retrieved 2026-08-16 |
| TaxEval v2Overall | 73.958 %Winner | 72.882 % | 2accuracyHigher is betterCohort: vals-tax-eval-v2:2:overallNo documented setup differenceWinner: GPT-5.4GPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-12Import ID: vals-tax_eval_v2-2026-08-16-5bbd495dac53Retrieved 2026-08-16 |
| Terminal-Bench 2.0Overall | 58.427 % | 67.416 %Winner | 2.0accuracyHigher is betterCohort: vals-terminal-bench-2:2.0:overallNo documented setup differenceWinner: Gemini 3.1 Pro PreviewGPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-06-04Import ID: vals-terminal-bench-2-2026-08-16-0d39f2e25723Retrieved 2026-08-16 |
| Vibe Code Bench v1.1Overall | 67.421 %Winner | 32.034 % | 1.1accuracyHigher is betterCohort: vals-vibe-code-bench:1.1:overallNo documented setup differenceWinner: GPT-5.4GPT-5.4: Snapshot: gpt-5.4-2026-03-05, Harness: OpenHands Gemini 3.1 Pro Preview: Harness: OpenHands | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-13Import ID: vals-vibe-code-2026-08-16-a73c9770bcc4Retrieved 2026-08-16 |
| VisualToolBenchOverall | 29.17% ± 0.13Higher measured value 95% confidence interval: ±0.13 | 28.97% ± 0.91 95% confidence interval: ±0.91 | 1,204-task releaseaverage pass rateHigher is betterCohort: scale-visual-toolbench-aprTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Attempts: pass@1 Gemini 3.1 Pro Preview: Attempts: pass@11,204 tasks, six tools, and at most 20 tool calls per task. | Scale AIType: Official benchmark leaderboardImport ID: curated-scale-ai-2026-08-05-8dae0c29c04e / curated-scale-ai-2026-08-05-18cb821a3540Retrieved 2026-08-05 |
| Arena Text, Style ControlOverall | 1,476.571 arena points 95% confidence interval: ±3.887The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result. | 1,486.245 arena pointsHigher measured value 95% confidence interval: ±3.337The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result. | style-control-v1style-controlled Bradley-Terry ratingHigher is betterCohort: lmarena-text-style-control:2026-08-12:overallTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control Gemini 3.1 Pro Preview: Snapshot: gemini-3.1-pro-preview, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 60,819 anonymous pairwise votes.Gemini 3.1 Pro Preview: Style-controlled Arena rating from 94,989 anonymous pairwise votes. | ArenaType: Official benchmark leaderboardLeaderboard updated: 2026-08-12Import ID: lmarena-eeedf9209373a7f9b324bfb839e7b02fedacd1e5-1611ed64f3f1Retrieved 2026-08-16 |
| Code MigrationCLI | 23.788 %Higher measured value | 10.833 % | 1accuracyHigher is betterCohort: vals-code-migration:1:cliTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-code-migration-2026-08-11-3a2d5ce718a3Retrieved 2026-08-11 |
| CorpFin v2Exact Pages | 64.103 %Higher measured value | 60.839 % | 2accuracyHigher is betterCohort: vals-corp-fin-v2:2:exact_pagesTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-06Import ID: vals-corp_fin_v2-2026-08-11-d9e58d191cd9Retrieved 2026-08-11 |
| CyberBenchPatch | 81.356 %Winner | 69.492 % | 1accuracyHigher is betterCohort: vals-cyberbench:1:patchNo documented setup differenceWinner: GPT-5.4GPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-03Import ID: vals-cyber-2026-08-11-1f20498a3dc1Retrieved 2026-08-11 |
| GPQA DiamondFew-Shot CoT | 91.414 % | 95.96 %Higher measured value | 1accuracyHigher is betterCohort: vals-gpqa:1:diamond_few_shot_cotTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,535 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-gpqa-2026-08-11-d3325290a404Retrieved 2026-08-11 |
| LegalBenchConclusion Tasks | 88.939 % | 89.257 %Higher measured value | 1accuracyHigher is betterCohort: vals-legal-bench:1:conclusion_tasksTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-legal_bench-2026-08-11-836e5e44618eRetrieved 2026-08-11 |
| LiveCodeBench v6Easy | 99.068 %No direct winner | 99.068 %No direct winner | 1accuracyHigher is betterCohort: vals-lcb:1:easyTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-lcb-2026-08-11-0342bcc364e4Retrieved 2026-08-11 |
| MedQAAsian | 96.15 % | 96.3 %Higher measured value | 1accuracyHigher is betterCohort: vals-medqa:1:asianTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-04-16Import ID: vals-medqa-2026-08-11-b50286439978Retrieved 2026-08-11 |
| MMLU-ProBiology | 92.887 % | 95.258 %Higher measured value | 1accuracyHigher is betterCohort: vals-mmlu-pro:1:biologyTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,535 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-mmlu_pro-2026-08-11-b8e53828d746Retrieved 2026-08-11 |
| MortgageTaxNumerical Extraction | 75.994 % | 76.391 %Higher measured value | 1accuracyHigher is betterCohort: vals-mortgage-tax:1:numerical_extractionTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-mortgage_tax-2026-08-11-5ea6eef3e482Retrieved 2026-08-11 |
| ProgramBenchAlmost Resolved | 7 %Higher measured value | 1 % | 1accuracyHigher is betterCohort: vals-programbench:1:almostTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-programbench-2026-08-11-b6f8959fbdfeRetrieved 2026-08-11 |
| SAGEAutomata | 21.428 % | 32.143 %Higher measured value | 1accuracyHigher is betterCohort: vals-sage:1:automataTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 30,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 30,000 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-sage-2026-08-11-f1f79f6d8af3Retrieved 2026-08-11 |
| SWE-bench Verified<15 min fix | 88.144 % | 88.66 %Higher measured value | 1accuracyHigher is betterCohort: vals-swebench:1:<15 min fixTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-08Import ID: vals-swebench-2026-08-11-a6212cd857d8Retrieved 2026-08-11 |
| TaxEval v2Correctness | 64.922 %Higher measured value | 60.749 % | 2accuracyHigher is betterCohort: vals-tax-eval-v2:2:correctnessTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,535 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-tax_eval_v2-2026-08-11-d05a40bf446fRetrieved 2026-08-11 |
| Terminal-Bench 2.0EasyReasoning: high | 100 %No direct winner | 100 %No direct winner | 2.0accuracyHigher is betterCohort: vals-terminal-bench-2:2.0:easyTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: high, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-06-04Import ID: vals-terminal-bench-2-2026-08-11-d91717be20e6Retrieved 2026-08-11 |
| Arena Text, Style ControlBusiness, management and finance | 1,474.218 arena points 95% confidence interval: ±6.56The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result. | 1,476.932 arena pointsNo clear lead 95% confidence interval: ±5.678The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result. | style-control-v1style-controlled Bradley-Terry ratingHigher is betterCohort: lmarena-text-style-control:2026-08-12:industry_business_and_management_and_financial_operationsNo documented setup differenceNo clear leadGPT-5.4: Snapshot: gpt-5.4, Harness: Arena text, style control Gemini 3.1 Pro Preview: Snapshot: gemini-3.1-pro-preview, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 13,072 anonymous pairwise votes.Gemini 3.1 Pro Preview: Style-controlled Arena rating from 18,759 anonymous pairwise votes. | ArenaType: Official benchmark leaderboardLeaderboard updated: 2026-08-12Import ID: lmarena-eeedf9209373a7f9b324bfb839e7b02fedacd1e5-1611ed64f3f1Retrieved 2026-08-16 |
| Code MigrationCOBOL | 68.571 %Higher measured value | 36.753 % | 1accuracyHigher is betterCohort: vals-code-migration:1:cobolTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-code-migration-2026-08-11-3a2d5ce718a3Retrieved 2026-08-11 |
| CorpFin v2Max Fitting Context | 64.918 % | 66.9 %Higher measured value | 2accuracyHigher is betterCohort: vals-corp-fin-v2:2:max_fitting_contextTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-06Import ID: vals-corp_fin_v2-2026-08-11-d9e58d191cd9Retrieved 2026-08-11 |
| CyberBenchPoC | 64.407 %Winner | 3.39 % | 1accuracyHigher is betterCohort: vals-cyberbench:1:pocNo documented setup differenceWinner: GPT-5.4GPT-5.4: Snapshot: gpt-5.4-2026-03-05 Gemini 3.1 Pro Preview: No further settings disclosed | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-03Import ID: vals-cyber-2026-08-11-1f20498a3dc1Retrieved 2026-08-11 |
| GPQA DiamondZero-Shot CoT | 91.919 % | 94.949 %Higher measured value | 1accuracyHigher is betterCohort: vals-gpqa:1:diamond_zero_shot_cotTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,535 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-gpqa-2026-08-11-d3325290a404Retrieved 2026-08-11 |
| LegalBenchInterpretation Tasks | 83.921 %Higher measured value | 83.823 % | 1accuracyHigher is betterCohort: vals-legal-bench:1:interpretation_tasksTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-legal_bench-2026-08-11-836e5e44618eRetrieved 2026-08-11 |
| LiveCodeBench v6Hard | 67.714 % | 76.571 %Higher measured value | 1accuracyHigher is betterCohort: vals-lcb:1:hardTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-lcb-2026-08-11-0342bcc364e4Retrieved 2026-08-11 |
| MedQABlack | 96.5 % | 96.8 %Higher measured value | 1accuracyHigher is betterCohort: vals-medqa:1:blackTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-04-16Import ID: vals-medqa-2026-08-11-b50286439978Retrieved 2026-08-11 |
| MMLU-ProBusiness | 91.762 % | 93.409 %Higher measured value | 1accuracyHigher is betterCohort: vals-mmlu-pro:1:businessTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,535 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-mmlu_pro-2026-08-11-b8e53828d746Retrieved 2026-08-11 |
| MortgageTaxSemantic Extraction | 60.652 % | 62.401 %Higher measured value | 1accuracyHigher is betterCohort: vals-mortgage-tax:1:semantic_extractionTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-mortgage_tax-2026-08-11-5ea6eef3e482Retrieved 2026-08-11 |
| ProgramBenchFully Resolved | 0 %No direct winner | 0 %No direct winner | 1accuracyHigher is betterCohort: vals-programbench:1:strictTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-programbench-2026-08-11-b6f8959fbdfeRetrieved 2026-08-11 |
| SAGECalculus | 70.988 % | 78.395 %Higher measured value | 1accuracyHigher is betterCohort: vals-sage:1:calculusTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 30,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 30,000 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-sage-2026-08-11-f1f79f6d8af3Retrieved 2026-08-11 |
| SWE-bench Verified>4 hours | 0 % | 33.333 %Higher measured value | 1accuracyHigher is betterCohort: vals-swebench:1:>4 hoursTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-08Import ID: vals-swebench-2026-08-11-a6212cd857d8Retrieved 2026-08-11 |
| TaxEval v2Stepwise Reasoning | 82.993 % | 85.016 %Higher measured value | 2accuracyHigher is betterCohort: vals-tax-eval-v2:2:stepwise_reasoningTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,535 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-tax_eval_v2-2026-08-11-d05a40bf446fRetrieved 2026-08-11 |
| Terminal-Bench 2.0HardReasoning: high | 33.333 % | 53.333 %Higher measured value | 2.0accuracyHigher is betterCohort: vals-terminal-bench-2:2.0:hardTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: high, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-06-04Import ID: vals-terminal-bench-2-2026-08-11-d91717be20e6Retrieved 2026-08-11 |
| Arena Text, Style ControlBusiness, management and finance | 1,483.617 arena pointsHigher measured value 95% confidence interval: ±6.683The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result. | 1,476.932 arena points 95% confidence interval: ±5.678The Arena score measures human preference rather than a fixed capability test. The confidence interval and model configuration are part of the result. | style-control-v1style-controlled Bradley-Terry ratingHigher is betterCohort: lmarena-text-style-control:2026-08-12:industry_business_and_management_and_financial_operationsTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-high, Reasoning: high, Harness: Arena text, style control Gemini 3.1 Pro Preview: Snapshot: gemini-3.1-pro-preview, Harness: Arena text, style controlGPT-5.4: Style-controlled Arena rating from 12,373 anonymous pairwise votes.Gemini 3.1 Pro Preview: Style-controlled Arena rating from 18,759 anonymous pairwise votes. | ArenaType: Official benchmark leaderboardLeaderboard updated: 2026-08-12Import ID: lmarena-eeedf9209373a7f9b324bfb839e7b02fedacd1e5-1611ed64f3f1Retrieved 2026-08-16 |
| Code MigrationCode Quality | 3.632 / 5Higher measured value | 2.538 / 5 | 1mean reviewer code-quality ratingHigher is betterCohort: vals-code-migration:1:code_qualityTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-code-migration-2026-08-11-3a2d5ce718a3Retrieved 2026-08-11 |
| CorpFin v2Shared Max Context | 66.783 %Higher measured value | 65.734 % | 2accuracyHigher is betterCohort: vals-corp-fin-v2:2:shared_max_contextTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-06Import ID: vals-corp_fin_v2-2026-08-11-d9e58d191cd9Retrieved 2026-08-11 |
| LegalBenchIssue Tasks | 90.526 % | 91.48 %Higher measured value | 1accuracyHigher is betterCohort: vals-legal-bench:1:issue_tasksTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-legal_bench-2026-08-11-836e5e44618eRetrieved 2026-08-11 |
| LiveCodeBench v6Medium | 85.64 % | 89.817 %Higher measured value | 1accuracyHigher is betterCohort: vals-lcb:1:mediumTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-lcb-2026-08-11-0342bcc364e4Retrieved 2026-08-11 |
| MedQAHispanic | 95.7 % | 95.95 %Higher measured value | 1accuracyHigher is betterCohort: vals-medqa:1:hispanicTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-04-16Import ID: vals-medqa-2026-08-11-b50286439978Retrieved 2026-08-11 |
| MMLU-ProChemistry | 91.078 % | 93.286 %Higher measured value | 1accuracyHigher is betterCohort: vals-mmlu-pro:1:chemistryTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,535 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-mmlu_pro-2026-08-11-b8e53828d746Retrieved 2026-08-11 |
| ProgramBenchRaw Pass Rate | 68.596 %Higher measured value | 39.468 % | 1accuracyHigher is betterCohort: vals-programbench:1:partialTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 128,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 65,536 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-programbench-2026-08-11-b6f8959fbdfeRetrieved 2026-08-11 |
| SAGELinear Algebra | 68.826 % | 74.205 %Higher measured value | 1accuracyHigher is betterCohort: vals-sage:1:linear_algebraTest setup not documented as identicalNo direct winnerGPT-5.4: Snapshot: gpt-5.4-2026-03-05, Reasoning: xhigh, Output limit: 30,000 tokens Gemini 3.1 Pro Preview: Reasoning: high, Temperature: 1, Output limit: 30,000 tokens | Vals AIType: Independent evaluationLeaderboard updated: 2026-08-09Import ID: vals-sage-2026-08-11-f1f79f6d8af3Retrieved 2026-08-11 |
The table shows 64 diversified result series from 138 shared results. This prevents one benchmark with many subtasks from crowding out every other test.
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.
Prices use official API rates per 1M tokens. Web subscriptions and cloud platform surcharges can differ.
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 Gemini 3.1 Pro Preview with other leading models using the same data and benchmark logic.
View the complete LLM comparison