Compare · ModelsLive · 2 picked · head to head
GPT-5.2 vs Kimi K2.5
Side by side · benchmarks, pricing, and signals you can act on.
Winner summary
GPT-5.2 wins on 13/16 benchmarks
GPT-5.2 wins 13 of 16 shared benchmarks. Leads in agentic · reasoning · knowledge.
Category leads
agentic·GPT-5.2reasoning·GPT-5.2knowledge·GPT-5.2general·Kimi K2.5math·GPT-5.2coding·GPT-5.2
Hype vs Reality
Attention vs performance
GPT-5.2
#154 by perf·#10 by attention
Kimi K2.5
#129 by perf·#18 by attention
Best value
Kimi K2.5
6.2x better value than GPT-5.2
GPT-5.2
6.0 pts/$
$7.88/M
Kimi K2.5
37.0 pts/$
$1.35/M
Vendor risk
Who is behind the model
OpenAI
$840.0B·Tier 1
Moonshot AI
$50.0B·Tier 1
Head to head
16 benchmarks · 2 models
GPT-5.2Kimi K2.5
APEX-Agents
GPT-5.2 leads by +19.9
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
Kimi K2.5
14.4
ARC-AGI
GPT-5.2 leads by +20.9
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
Kimi K2.5
65.3
ARC-AGI-2
GPT-5.2 leads by +41.1
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
Kimi K2.5
11.8
Chess Puzzles
GPT-5.2 leads by +37.9
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
Kimi K2.5
7.4
Cl Bench
Kimi K2.5 leads by +1.1
GPT-5.2
18.2
Kimi K2.5
19.3
FrontierMath-2025-02-28-Private
Kimi K2.5 leads by +8.3
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
Kimi K2.5
49.0
FrontierMath-Tier-4-2025-07-01-Private
GPT-5.2 leads by +11.8
FrontierMath Tier 4 (Jul 2025) · the most challenging tier of frontier mathematics, containing problems that push the absolute limits of AI mathematical reasoning.
GPT-5.2
18.8
Kimi K2.5
7.0
GPQA diamond
GPT-5.2 leads by +5.1
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
88.5
Kimi K2.5
83.5
HLE
GPT-5.2 leads by +3.6
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-5.2
24.2
Kimi K2.5
20.6
OTIS Mock AIME 2024-2025
GPT-5.2 leads by +3.9
OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills.
GPT-5.2
96.1
Kimi K2.5
92.2
PostTrainBench
GPT-5.2 leads by +11.1
GPT-5.2
21.4
Kimi K2.5
10.3
SimpleBench
Kimi K2.5 leads by +1.2
SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking.
GPT-5.2
35.0
Kimi K2.5
36.2
SimpleQA Verified
GPT-5.2 leads by +2.8
SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information.
GPT-5.2
37.1
Kimi K2.5
34.3
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.
GPT-5.2
73.8
Kimi K2.5
73.8
Terminal Bench
GPT-5.2 leads by +21.7
Terminal-Bench 2.0 · evaluates AI agents on real terminal-based coding tasks · writing scripts, debugging, running tests, and managing projects entirely through command-line interaction. Tests both code quality and terminal fluency. Claude Opus 4.7 scores 69.4%, demonstrating significant agentic terminal competence.
GPT-5.2
64.9
Kimi K2.5
43.2
WeirdML
GPT-5.2 leads by +26.6
WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns.
GPT-5.2
72.2
Kimi K2.5
45.6
Full benchmark table
| Benchmark | GPT-5.2 | Kimi K2.5 |
|---|---|---|
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 | 14.4 |
ARC-AGI ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization. | 86.2 | 65.3 |
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 | 11.8 |
Chess Puzzles Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities. | 45.3 | 7.4 |
Cl Bench | 18.2 | 19.3 |
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 | 49.0 |
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. | 18.8 | 7.0 |
GPQA diamond Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs. | 88.5 | 83.5 |
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%. | 24.2 | 20.6 |
OTIS Mock AIME 2024-2025 OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills. | 96.1 | 92.2 |
PostTrainBench | 21.4 | 10.3 |
SimpleBench SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking. | 35.0 | 36.2 |
SimpleQA Verified SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information. | 37.1 | 34.3 |
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. | 73.8 | 73.8 |
Terminal Bench Terminal-Bench 2.0 · evaluates AI agents on real terminal-based coding tasks · writing scripts, debugging, running tests, and managing projects entirely through command-line interaction. Tests both code quality and terminal fluency. Claude Opus 4.7 scores 69.4%, demonstrating significant agentic terminal competence. | 64.9 | 43.2 |
WeirdML WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns. | 72.2 | 45.6 |
Pricing · per 1M tokens · projected $/mo at 10M tokens