Compare · ModelsLive · 2 picked · head to head
GPT-5.2 vs o3
Side by side · benchmarks, pricing, and signals you can act on.
Winner summary
GPT-5.2 wins on 20/24 benchmarks
GPT-5.2 wins 20 of 24 shared benchmarks. Leads in agentic · reasoning · knowledge.
Category leads
agentic·GPT-5.2reasoning·GPT-5.2knowledge·GPT-5.2general·GPT-5.2math·GPT-5.2coding·GPT-5.2
Hype vs Reality
Attention vs performance
GPT-5.2
#168 by perf·#9 by attention
o3
#121 by perf·no signal
Best value
o3
1.8x better value than GPT-5.2
GPT-5.2
5.8 pts/$
$7.88/M
o3
10.3 pts/$
$5.00/M
Vendor risk
Who is behind the model
OpenAI
$840.0B·Tier 1
OpenAI
$840.0B·Tier 1
Head to head
24 benchmarks · 2 models
GPT-5.2o3
APEX-Agents
GPT-5.2 leads by +17.1
APEX-Agents · evaluates AI agents on complex, multi-step tasks requiring planning, tool use, and autonomous decision-making in realistic environments.
GPT-5.2
34.3
o3
17.2
ARC-AGI
GPT-5.2 leads by +25.4
ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization.
GPT-5.2
86.2
o3
60.8
ARC-AGI-2
GPT-5.2 leads by +46.4
ARC-AGI-2 · the second iteration of the Abstraction and Reasoning Corpus, testing novel pattern recognition and abstract reasoning without prior training data.
GPT-5.2
52.9
o3
6.5
Chess Puzzles
GPT-5.2 leads by +11.6
Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities.
GPT-5.2
45.3
o3
33.7
Cl Bench
GPT-5.2 leads by +0.4
GPT-5.2
18.2
o3
17.8
DeepResearch Bench
o3 leads by +4.1
DeepResearch Bench · evaluates AI on complex multi-step research tasks requiring information gathering, synthesis, and producing comprehensive analyses.
GPT-5.2
41.1
o3
45.2
Dtbench
GPT-5.2 leads by +10.2
GPT-5.2
84.9
o3
74.7
FrontierMath-2025-02-28-Private
GPT-5.2 leads by +22.0
FrontierMath (Feb 2025) · original research-level math problems created by mathematicians, testing capabilities at the boundary of current AI mathematical reasoning.
GPT-5.2
40.7
o3
18.7
FrontierMath-Tier-4-2025-07-01-Private
GPT-5.2 leads by +15.3
FrontierMath Tier 4 (Jul 2025) · the most challenging tier of frontier mathematics, containing problems that push the absolute limits of AI mathematical reasoning.
GPT-5.2
18.8
o3
3.5
FrontierMath-Tiers-1-3-v2-Private
GPT-5.2 leads by +35.9
GPT-5.2
71.0
o3
35.1
Gdpval
GPT-5.2 leads by +18.9
GPT-5.2
49.7
o3
30.8
GPQA diamond
GPT-5.2 leads by +16.5
Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs.
GPT-5.2
92.2
o3
75.8
GSO-Bench
GPT-5.2 leads by +18.6
GSO-Bench · evaluates AI models on real-world open-source software engineering tasks, testing the ability to understand and resolve actual GitHub issues.
GPT-5.2
27.4
o3
8.8
HLE
GPT-5.2 leads by +7.9
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.
GPT-5.2
24.2
o3
16.3
Lmca
GPT-5.2 leads by +4.9
GPT-5.2
51.7
o3
46.7
Metr Time Horizons
GPT-5.2 leads by +9.9
GPT-5.2
75.3
o3
65.4
Mystery Game Puzzles
o3 leads by +6.6
GPT-5.2
15.2
o3
21.8
OTIS Mock AIME 2024-2025
GPT-5.2 leads by +11.7
OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills.
GPT-5.2
96.1
o3
84.4
SimpleBench
o3 leads by +8.8
SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking.
GPT-5.2
35.0
o3
43.7
SimpleQA Verified
o3 leads by +14.1
SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information.
GPT-5.2
38.9
o3
53.0
SWE-Bench verified
GPT-5.2 leads by +11.4
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. The benchmark remains the most-cited evaluation for code-generation capability.
GPT-5.2
73.8
o3
62.3
SWE-Bench Verified (Bash Only)
GPT-5.2 leads by +13.4
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-5.2
71.8
o3
58.4
VPCT
GPT-5.2 leads by +48.0
VPCT (Visual Pattern Completion Test) · tests visual reasoning and pattern recognition by having models complete visual sequences and transformations.
GPT-5.2
76.0
o3
28.0
WeirdML
GPT-5.2 leads by +19.8
WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns.
GPT-5.2
72.2
o3
52.4
Full benchmark table
| Benchmark | GPT-5.2 | o3 |
|---|---|---|
APEX-Agents APEX-Agents · evaluates AI agents on complex, multi-step tasks requiring planning, tool use, and autonomous decision-making in realistic environments. | 34.3 | 17.2 |
ARC-AGI ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization. | 86.2 | 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. | 52.9 | 6.5 |
Chess Puzzles Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities. | 45.3 | 33.7 |
Cl Bench | 18.2 | 17.8 |
DeepResearch Bench DeepResearch Bench · evaluates AI on complex multi-step research tasks requiring information gathering, synthesis, and producing comprehensive analyses. | 41.1 | 45.2 |
Dtbench | 84.9 | 74.7 |
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. | 40.7 | 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. | 18.8 | 3.5 |
FrontierMath-Tiers-1-3-v2-Private | 71.0 | 35.1 |
Gdpval | 49.7 | 30.8 |
GPQA diamond Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs. | 92.2 | 75.8 |
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. | 27.4 | 8.8 |
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. | 24.2 | 16.3 |
Lmca | 51.7 | 46.7 |
Metr Time Horizons | 75.3 | 65.4 |
Mystery Game Puzzles | 15.2 | 21.8 |
OTIS Mock AIME 2024-2025 OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills. | 96.1 | 84.4 |
SimpleBench SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking. | 35.0 | 43.7 |
SimpleQA Verified SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information. | 38.9 | 53.0 |
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. The benchmark remains the most-cited evaluation for code-generation capability. | 73.8 | 62.3 |
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. | 71.8 | 58.4 |
VPCT VPCT (Visual Pattern Completion Test) · tests visual reasoning and pattern recognition by having models complete visual sequences and transformations. | 76.0 | 28.0 |
WeirdML WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns. | 72.2 | 52.4 |
Pricing · per 1M tokens · projected $/mo at 10M tokens