GPT-5.2_Thinking_A_New_Benchmark
1
2
3
4
5
6
7
8
9
10
11
Slide 1
NotebookLM to PPTXTurn NotebookLM Slides into Editable PPTXExport PPTX
Log In
NotebookLM
OpenAl models were run with maximum available reasoning effort.
A New State-of-the-Art in Al Reasoning
Introducing GPT-5.2 Thinking
Software Engineering
55.6%
Abstract Reasoning
86.2%
NotebookLM
Science
92.4%
GPT-5.2 Thinking significantly outperforms all current competitor models and its direct predecessors across a wide range of difficult reasoning tasks. This deck explores the key domains where it sets a new standard for Al capability.
Mathematics
100.0%
GPT-5.2 Thinking Establishes a New Performance Frontier Across Major Al Benchmarks.
NotebookLM
Claude Opus 4.5
GPT-5.1 Thinking
Gemini 3 Pro
GPT-5.2 Thinking
90%
92%
problem-solving on par with the American Invitational Mathematics Examination, a highly challenging competition for pre-college students. Solved without tool use.
Evaluates mathematical
AIME 2025:
94%
96%
98%
92.8%
94.0%
95.0%
100.0%
GPT-5.2 Achieves a Perfect Score on Competition-Level Mathematics.
100%
NotebookLM
mathematics problems (Tier 1-3, Tier 4) testing deep reasoning capabilities.
A benchmark of advanced
FrontierMath:
50%
40%
30%
20%
10%
0%
Leading the next best model by 2.7 percentage points.
Claude Opus 4.5
31.0%
GPT-5.1 Thinking
37.6%
Gemini 3 Pro
40.3%
GPT-5.2 Thinking
The Model Pushes the Boundaries of Advanced Mathematical Reasoning.
NotebookLM
Claude Opus 4.5
GPT-5.1 Thinking
Gemini 3 Pro
GPT-5.2 Thinking
CharXiv Reasoning: Tests the ability to answer questions based on scientific figures and diagrams from research papers.
GPQA Diamond: A set of challenging science questions that stump even domain experts.
CharXiv Reasoning
GPQA Diamond
67.0%
81.4%
82.1%
91.9%
92.4%
87.0%
88.1%
GPT-5.2 Demonstrates Superiority in Nuanced Scientific Inquiry.
Reasoning Corpus (ARC) is a difficult benchmark designed to test AGl-like fluid intelligence and abstract reasoning on novel problems.
The Abstraction and
ARC-AGI:
NotebookLM
GPT-5.1 Thinking
Gemini 3 Pro
Claude Opus 4.5
GPT-5.2 Thinking
0%
20%
40%
52.9%
37.6%
17.6%
31.1%
60%
A Fundamental Leap in Abstract Reasoning Signals Progress on Core Al Challenges
NotebookLM
The leap in ARC-AGl 2 performance from 17.6% to 52.9% represents a monumental step forward, demonstrating a rapidly advancing ability to solve novel problems that require true abstract thought
GPT-5.2 Thinking
GPT-5.1 Thinking
17.6%
Improvement
2.9x
52.9%
Abstract Reasoning Capability Increases Nearly 3x Over the Previous Generation.
NotebookLM
software engineering tasks sourced from GitHub issues, a direct measure of practical coding and problem-solving.
Tests the ability to autonomously resolve real-world
SWE-Bench Pro:
Gemini 3 Pro
GPT-5.1 Thinking
Claude Opus 4.5
GPT-5.2 Thinking
0%
50.8%
10%
20%
30%
40%
50%
52.0%
43.3%
55.6%
60%
SWE-Bench Pro Scores
The Model Sets a New Standard for Autonomous Software Engineering.
NotebookLM
GDPval: Evaluates performance across a diverse set of graduate-level professional knowledge work tasks, simulating complex real-world assignments.
4.5
GPT-5
Gemini 3 Pro
GDPval GPT-5.2 Thinking Claude Opus
38.8%
53.5%
59.6%
83% Improvement
70.9%
Performance on Complex Knowledge Work Tasks Nearly Doubles in One Generation.
NotebookLM
"CpenAl models were run with maximum available reasoning effort.
48.8%
42.4%
59.9%
36.9%
83.0%
72.0%
88.9%
GPT-5.1 Thinking
43.0%
35.3%
72.9%
53.0%
89.9%
76.9%
96.5%
Claude Opus 4.5
59.9%
48.5%
28.5%
53.4%
80.2%
3.6%
98.8%
Gemini 3 Pro
ARC-AGI2
GDPval
SWE-Bench Pro
GPQA Diamond
FrontierMath
AIME 2025
Model
46.9%
43.2%
68.3%
86.3%
98.6%
63.7%
96.5%
GPT-5.2Thinking
CharXiv Reasoning
GPT-5.1 Thinking
Claude Opus 4.5
Gemini 3 Pro
GPT-5.2 Thinking
SWE-Bench Pro
96.9%
CharXiv Reasoning
GDPval
58.1%
58.2%
90.3%
57.3%
82%
53.2%
S7%
8076
GPQA Diamond
93%
72.7%
38%
92%
80%
98%
ARC-AGI1
57%
55%
95.5%
99%
58.1%
98.3%
98.8%
FrontierMath
80.2%
ARC-AGI2
97.6%
AIME 2025
GPT-5.2 Thinking Consistently Outperforms Across Every Benchmark
Breakthrough Abstract Reasoning: Showed a near 3x generational leap in solving novel abstract problems.
Flawless Mathematical Execution: Achieved a perfect 100%on competition- level mathematics.
GPT-5.2 Thinking's performance redefines the state-of-the-art in artificial intelligence. Its capabilities demonstrate a new level of proficiency in areas critical for scientific and technological progress.
A New Benchmark for Intelligence
NotebookLM
OpenAI
Dominance in Applied Tasks: Established new leadership in complex domains like software engineering and scientific analysis.