Compare · ModelsLive · 3 picked · head to head
GPT-5.4 Pro vs Gemini 3.1 Pro Preview vs Gemini 3.5 Flash
Side by side · benchmarks, pricing, and signals you can act on.
Winner summary
Gemini 3.1 Pro Preview wins on 8/20 benchmarks
Gemini 3.1 Pro Preview wins 8 of 20 shared benchmarks. Leads in reasoning · coding.
Category leads
reasoning·Gemini 3.1 Pro Previewknowledge·GPT-5.4 Promath·GPT-5.4 Procoding·Gemini 3.1 Pro Previewspeed·Gemini 3.5 Flashagentic·Gemini 3.5 Flasharena·Gemini 3.5 Flash
Hype vs Reality
Attention vs performance
GPT-5.4 Pro
#41 by perf·no signal
Gemini 3.1 Pro Preview
#74 by perf·no signal
Gemini 3.5 Flash
#44 by perf·no signal
Best value
Gemini 3.5 Flash
1.5x better value than Gemini 3.1 Pro Preview
GPT-5.4 Pro
0.6 pts/$
$105.00/M
Gemini 3.1 Pro Preview
8.1 pts/$
$7.00/M
Gemini 3.5 Flash
11.9 pts/$
$5.25/M
Vendor risk
Who is behind the model
OpenAI
$840.0B·Tier 1
Google DeepMind
$4.00T·Tier 1
Google DeepMind
$4.00T·Tier 1
Head to head
20 benchmarks · 3 models
GPT-5.4 ProGemini 3.1 Pro PreviewGemini 3.5 Flash
ARC-AGI
Gemini 3.1 Pro Preview leads by +3.5
ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization.
GPT-5.4 Pro
94.5
Gemini 3.1 Pro Preview
98.0
Gemini 3.5 Flash
92.5
ARC-AGI-2
GPT-5.4 Pro leads by +6.2
ARC-AGI-2 · the second iteration of the Abstraction and Reasoning Corpus, testing novel pattern recognition and abstract reasoning without prior training data.
GPT-5.4 Pro
83.3
Gemini 3.1 Pro Preview
77.1
Gemini 3.5 Flash
72.1
Chess Puzzles
GPT-5.4 Pro leads by +3.6
Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities.
GPT-5.4 Pro
58.6
Gemini 3.1 Pro Preview
55.0
Gemini 3.5 Flash
50.0
FrontierMath-2025-02-28-Private
GPT-5.4 Pro leads by +11.0
FrontierMath (Feb 2025) · original research-level math problems created by mathematicians, testing capabilities at the boundary of current AI mathematical reasoning.
GPT-5.4 Pro
50.0
Gemini 3.1 Pro Preview
36.9
Gemini 3.5 Flash
39.0
FrontierMath-Tier-4-2025-07-01-Private
GPT-5.4 Pro leads by +20.8
FrontierMath Tier 4 (Jul 2025) · the most challenging tier of frontier mathematics, containing problems that push the absolute limits of AI mathematical reasoning.
GPT-5.4 Pro
37.5
Gemini 3.1 Pro Preview
16.7
Gemini 3.5 Flash
14.6
GPQA diamond
GPT-5.4 Pro leads by +0.7
Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs.
GPT-5.4 Pro
92.8
Gemini 3.1 Pro Preview
92.1
Gemini 3.5 Flash
90.4
SimpleBench
Gemini 3.1 Pro Preview leads by +3.5
SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking.
GPT-5.4 Pro
68.9
Gemini 3.1 Pro Preview
75.5
Gemini 3.5 Flash
72.0
SimpleQA Verified
Gemini 3.1 Pro Preview leads by +8.9
SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information.
GPT-5.4 Pro
47.8
Gemini 3.1 Pro Preview
77.3
Gemini 3.5 Flash
68.4
WeirdML
Gemini 3.1 Pro Preview leads by +9.5
WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns.
GPT-5.4 Pro
57.4
Gemini 3.1 Pro Preview
72.1
Gemini 3.5 Flash
62.6
Artificial Analysis · Agentic Index
Gemini 3.5 Flash leads by +16.1
Artificial Analysis Agentic Index · a composite score measuring how well a model performs in agentic workflows · multi-step tool use, planning, error recovery, and autonomous task completion. Aggregates results from multiple agentic benchmarks including SWE-bench, tool-use tests, and planning evaluations. The canonical single-number metric for "how good is this model as an agent?"
Gemini 3.1 Pro Preview
21.4
Gemini 3.5 Flash
37.5
Artificial Analysis · Coding Index
Gemini 3.5 Flash leads by +1.3
Artificial Analysis Coding Index · a composite score that aggregates performance across multiple coding benchmarks into a single index. Tracks code generation quality, debugging ability, multi-language competence, and real-world software engineering tasks. Used by Artificial Analysis to rank model coding capability in a normalized, comparable format. Useful for developers choosing between models for coding-heavy workloads.
Gemini 3.1 Pro Preview
68.8
Gemini 3.5 Flash
70.1
Artificial Analysis · Quality Index
Gemini 3.5 Flash leads by +3.7
Gemini 3.1 Pro Preview
46.5
Gemini 3.5 Flash
50.2
APEX-Agents
Gemini 3.5 Flash leads by +16.1
APEX-Agents · evaluates AI agents on complex, multi-step tasks requiring planning, tool use, and autonomous decision-making in realistic environments.
Gemini 3.1 Pro Preview
33.5
Gemini 3.5 Flash
49.6
Chatbot Arena Elo · Coding
Gemini 3.5 Flash leads by +58.6
Gemini 3.1 Pro Preview
1447.0
Gemini 3.5 Flash
1505.5
Chatbot Arena Elo · Overall
Gemini 3.1 Pro Preview leads by +10.2
Gemini 3.1 Pro Preview
1486.4
Gemini 3.5 Flash
1476.3
FrontierMath-Tier-4-v2-Private
Gemini 3.1 Pro Preview
26.8
Gemini 3.5 Flash
26.8
FrontierMath-Tiers-1-3-v2-Private
Gemini 3.5 Flash leads by +3.2
Gemini 3.1 Pro Preview
59.6
Gemini 3.5 Flash
62.8
HLE
Gemini 3.1 Pro Preview leads by +2.2
HLE (Humanity's Last Exam) · a reasoning benchmark designed to be the hardest public evaluation of AI. Questions span mathematics, physics, philosophy, and logic · curated to be at or beyond the frontier of human expert capability. Tested with and without tool augmentation. Claude Opus 4.7 scores 46.9% without tools and 54.7% with tools · making it one of the few benchmarks where the top score is below 60%.
GPT-5.4 Pro
41.5
Gemini 3.1 Pro Preview
43.7
OTIS Mock AIME 2024-2025
Gemini 3.1 Pro Preview leads by +0.0
OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills.
Gemini 3.1 Pro Preview
95.6
Gemini 3.5 Flash
95.5
SWE-Bench verified
Gemini 3.5 Flash leads by +3.7
SWE-bench Verified · 500 human-validated tasks from 12 real Python repositories (Django, Flask, scikit-learn, sympy, and others). Each task requires the model to produce a git patch that resolves a real GitHub issue and passes the test suite. The verified subset eliminates ambiguous tasks from the original SWE-bench. Claude Mythos Preview leads at 93.9%, crossing 90% for the first time in 2026. Opus 4.6 scores 80.8%. The benchmark remains the most-cited evaluation for code-generation capability.
Gemini 3.1 Pro Preview
75.6
Gemini 3.5 Flash
79.3
Full benchmark table
| Benchmark | GPT-5.4 Pro | Gemini 3.1 Pro Preview | Gemini 3.5 Flash |
|---|---|---|---|
ARC-AGI ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization. | 94.5 | 98.0 | 92.5 |
ARC-AGI-2 ARC-AGI-2 · the second iteration of the Abstraction and Reasoning Corpus, testing novel pattern recognition and abstract reasoning without prior training data. | 83.3 | 77.1 | 72.1 |
Chess Puzzles Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities. | 58.6 | 55.0 | 50.0 |
FrontierMath-2025-02-28-Private FrontierMath (Feb 2025) · original research-level math problems created by mathematicians, testing capabilities at the boundary of current AI mathematical reasoning. | 50.0 | 36.9 | 39.0 |
FrontierMath-Tier-4-2025-07-01-Private FrontierMath Tier 4 (Jul 2025) · the most challenging tier of frontier mathematics, containing problems that push the absolute limits of AI mathematical reasoning. | 37.5 | 16.7 | 14.6 |
GPQA diamond Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs. | 92.8 | 92.1 | 90.4 |
SimpleBench SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking. | 68.9 | 75.5 | 72.0 |
SimpleQA Verified SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information. | 47.8 | 77.3 | 68.4 |
WeirdML WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns. | 57.4 | 72.1 | 62.6 |
Artificial Analysis · Agentic Index Artificial Analysis Agentic Index · a composite score measuring how well a model performs in agentic workflows · multi-step tool use, planning, error recovery, and autonomous task completion. Aggregates results from multiple agentic benchmarks including SWE-bench, tool-use tests, and planning evaluations. The canonical single-number metric for "how good is this model as an agent?" | — | 21.4 | 37.5 |
Artificial Analysis · Coding Index Artificial Analysis Coding Index · a composite score that aggregates performance across multiple coding benchmarks into a single index. Tracks code generation quality, debugging ability, multi-language competence, and real-world software engineering tasks. Used by Artificial Analysis to rank model coding capability in a normalized, comparable format. Useful for developers choosing between models for coding-heavy workloads. | — | 68.8 | 70.1 |
Artificial Analysis · Quality Index | — | 46.5 | 50.2 |
APEX-Agents APEX-Agents · evaluates AI agents on complex, multi-step tasks requiring planning, tool use, and autonomous decision-making in realistic environments. | — | 33.5 | 49.6 |
Chatbot Arena Elo · Coding | — | 1447.0 | 1505.5 |
Chatbot Arena Elo · Overall | — | 1486.4 | 1476.3 |
FrontierMath-Tier-4-v2-Private | — | 26.8 | 26.8 |
FrontierMath-Tiers-1-3-v2-Private | — | 59.6 | 62.8 |
HLE HLE (Humanity's Last Exam) · a reasoning benchmark designed to be the hardest public evaluation of AI. Questions span mathematics, physics, philosophy, and logic · curated to be at or beyond the frontier of human expert capability. Tested with and without tool augmentation. Claude Opus 4.7 scores 46.9% without tools and 54.7% with tools · making it one of the few benchmarks where the top score is below 60%. | 41.5 | 43.7 | — |
OTIS Mock AIME 2024-2025 OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills. | — | 95.6 | 95.5 |
SWE-Bench verified SWE-bench Verified · 500 human-validated tasks from 12 real Python repositories (Django, Flask, scikit-learn, sympy, and others). Each task requires the model to produce a git patch that resolves a real GitHub issue and passes the test suite. The verified subset eliminates ambiguous tasks from the original SWE-bench. Claude Mythos Preview leads at 93.9%, crossing 90% for the first time in 2026. Opus 4.6 scores 80.8%. The benchmark remains the most-cited evaluation for code-generation capability. | — | 75.6 | 79.3 |
Pricing · per 1M tokens · projected $/mo at 10M tokens
| Model | Input | Output | Context | Projected $/mo |
|---|---|---|---|---|
| $30.00 | $180.00 | 1.1M tokens (~525 books) | $675.00 | |
| $2.00 | $12.00 | 1.0M tokens (~524 books) | $45.00 | |
| $1.50 | $9.00 | 1.0M tokens (~524 books) | $33.75 |
People also compared