Compare · ModelsLive · 3 picked · head to head
GPT-5 Mini vs o4 Mini vs o3
Side by side · benchmarks, pricing, and signals you can act on.
Winner summary
o3 wins on 14/25 benchmarks
o3 wins 14 of 25 shared benchmarks. Leads in reasoning · knowledge.
Category leads
reasoning·o3knowledge·o3math·GPT-5 Minilanguage·o4 Minicoding·GPT-5 Mini
Hype vs Reality
Attention vs performance
GPT-5 Mini
#65 by perf·no signal
o4 Mini
#81 by perf·#13 by attention
o3
#69 by perf·no signal
Best value
GPT-5 Mini
2.6x better value than o4 Mini
GPT-5 Mini
49.8 pts/$
$1.13/M
o4 Mini
19.3 pts/$
$2.75/M
o3
11.0 pts/$
$5.00/M
Vendor risk
Who is behind the model
OpenAI
$840.0B·Tier 1
OpenAI
$840.0B·Tier 1
OpenAI
$840.0B·Tier 1
Head to head
25 benchmarks · 3 models
GPT-5 Minio4 Minio3
ARC-AGI
o3 leads by +2.1
ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization.
GPT-5 Mini
54.3
o4 Mini
58.7
o3
60.8
ARC-AGI-2
o3 leads by +0.4
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 Mini
4.4
o4 Mini
6.1
o3
6.5
Fiction.LiveBench
o3 leads by +11.1
Fiction.LiveBench · a continuously updated benchmark using recently published fiction to test reading comprehension and reasoning, preventing data contamination.
GPT-5 Mini
69.4
o4 Mini
77.8
o3
88.9
FrontierMath-2025-02-28-Private
GPT-5 Mini leads by +2.4
FrontierMath (Feb 2025) · original research-level math problems created by mathematicians, testing capabilities at the boundary of current AI mathematical reasoning.
GPT-5 Mini
27.2
o4 Mini
24.8
o3
18.7
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.
GPT-5 Mini
6.3
o4 Mini
6.3
o3
2.1
GPQA diamond
o3 leads by +3.0
Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs.
GPT-5 Mini
66.7
o4 Mini
72.8
o3
75.8
HELM · GPQA
GPT-5 Mini leads by +0.3
GPT-5 Mini
75.6
o4 Mini
73.5
o3
75.3
HELM · IFEval
o4 Mini leads by +0.2
GPT-5 Mini
92.7
o4 Mini
92.9
o3
86.9
HELM · MMLU-Pro
o3 leads by +2.4
GPT-5 Mini
83.5
o4 Mini
82.0
o3
85.9
HELM · Omni-MATH
GPT-5 Mini leads by +0.2
GPT-5 Mini
72.2
o4 Mini
72.0
o3
71.4
HELM · WildBench
o3 leads by +0.6
GPT-5 Mini
85.5
o4 Mini
85.4
o3
86.1
HLE
o3 leads by +0.9
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 Mini
15.4
o4 Mini
13.9
o3
16.3
MATH level 5
GPT-5 Mini leads by +0.0
MATH Level 5 · the hardest tier of the MATH benchmark, featuring competition-level problems from AMC, AIME, and Olympiad-style mathematics.
GPT-5 Mini
97.8
o4 Mini
97.8
o3
97.8
OTIS Mock AIME 2024-2025
GPT-5 Mini leads by +2.8
OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills.
GPT-5 Mini
86.7
o4 Mini
81.7
o3
83.9
SimpleQA Verified
o3 leads by +29.1
SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information.
GPT-5 Mini
21.0
o4 Mini
23.9
o3
53.0
SWE-Bench Verified (Bash Only)
GPT-5 Mini leads by +1.4
SWE-Bench Verified (Bash Only) · a curated subset of SWE-bench where models fix real Python repository bugs using only bash commands, no agent frameworks.
GPT-5 Mini
59.8
o4 Mini
45.0
o3
58.4
VPCT
o4 Mini leads by +8.3
VPCT (Visual Pattern Completion Test) · tests visual reasoning and pattern recognition by having models complete visual sequences and transformations.
GPT-5 Mini
10.3
o4 Mini
36.3
o3
28.0
WeirdML
GPT-5 Mini leads by +0.1
WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns.
GPT-5 Mini
52.7
o4 Mini
52.6
o3
52.4
Aider polyglot
o3 leads by +9.3
Aider Polyglot · measures how well AI models can edit code across multiple programming languages using the Aider coding assistant framework.
o4 Mini
72.0
o3
81.3
CadEval
o3 leads by +12.0
CadEval · evaluates the ability to generate and reason about Computer-Aided Design code, testing spatial reasoning and engineering knowledge.
o4 Mini
62.0
o3
74.0
GeoBench
o3 leads by +10.0
GeoBench · tests geographic knowledge and spatial reasoning across countries, landmarks, coordinates, and geopolitical understanding.
o4 Mini
64.0
o3
74.0
GSO-Bench
o3 leads by +5.2
GSO-Bench · evaluates AI models on real-world open-source software engineering tasks, testing the ability to understand and resolve actual GitHub issues.
o4 Mini
3.6
o3
8.8
Lech Mazur Writing
o3 leads by +8.9
Lech Mazur Writing · evaluates creative writing ability, assessing prose quality, narrative coherence, and stylistic sophistication.
o4 Mini
75.0
o3
83.9
SimpleBench
o3 leads by +17.3
SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking.
o4 Mini
26.4
o3
43.7
SWE-Bench verified
GPT-5 Mini leads by +2.4
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.
GPT-5 Mini
64.7
o3
62.3
Full benchmark table
| Benchmark | GPT-5 Mini | o4 Mini | o3 |
|---|---|---|---|
ARC-AGI ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization. | 54.3 | 58.7 | 60.8 |
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. | 4.4 | 6.1 | 6.5 |
Fiction.LiveBench Fiction.LiveBench · a continuously updated benchmark using recently published fiction to test reading comprehension and reasoning, preventing data contamination. | 69.4 | 77.8 | 88.9 |
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. | 27.2 | 24.8 | 18.7 |
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. | 6.3 | 6.3 | 2.1 |
GPQA diamond Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs. | 66.7 | 72.8 | 75.8 |
HELM · GPQA | 75.6 | 73.5 | 75.3 |
HELM · IFEval | 92.7 | 92.9 | 86.9 |
HELM · MMLU-Pro | 83.5 | 82.0 | 85.9 |
HELM · Omni-MATH | 72.2 | 72.0 | 71.4 |
HELM · WildBench | 85.5 | 85.4 | 86.1 |
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%. | 15.4 | 13.9 | 16.3 |
MATH level 5 MATH Level 5 · the hardest tier of the MATH benchmark, featuring competition-level problems from AMC, AIME, and Olympiad-style mathematics. | 97.8 | 97.8 | 97.8 |
OTIS Mock AIME 2024-2025 OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills. | 86.7 | 81.7 | 83.9 |
SimpleQA Verified SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information. | 21.0 | 23.9 | 53.0 |
SWE-Bench Verified (Bash Only) SWE-Bench Verified (Bash Only) · a curated subset of SWE-bench where models fix real Python repository bugs using only bash commands, no agent frameworks. | 59.8 | 45.0 | 58.4 |
VPCT VPCT (Visual Pattern Completion Test) · tests visual reasoning and pattern recognition by having models complete visual sequences and transformations. | 10.3 | 36.3 | 28.0 |
WeirdML WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns. | 52.7 | 52.6 | 52.4 |
Aider polyglot Aider Polyglot · measures how well AI models can edit code across multiple programming languages using the Aider coding assistant framework. | — | 72.0 | 81.3 |
CadEval CadEval · evaluates the ability to generate and reason about Computer-Aided Design code, testing spatial reasoning and engineering knowledge. | — | 62.0 | 74.0 |
GeoBench GeoBench · tests geographic knowledge and spatial reasoning across countries, landmarks, coordinates, and geopolitical understanding. | — | 64.0 | 74.0 |
GSO-Bench GSO-Bench · evaluates AI models on real-world open-source software engineering tasks, testing the ability to understand and resolve actual GitHub issues. | — | 3.6 | 8.8 |
Lech Mazur Writing Lech Mazur Writing · evaluates creative writing ability, assessing prose quality, narrative coherence, and stylistic sophistication. | — | 75.0 | 83.9 |
SimpleBench SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking. | — | 26.4 | 43.7 |
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. | 64.7 | — | 62.3 |
Pricing · per 1M tokens · projected $/mo at 10M tokens
| Model | Input | Output | Context | Projected $/mo |
|---|---|---|---|---|
| $0.25 | $2.00 | 400K tokens (~200 books) | $6.88 | |
| $1.10 | $4.40 | 200K tokens (~100 books) | $19.25 | |
| $2.00 | $8.00 | 200K tokens (~100 books) | $35.00 |