Compare · ModelsLive · 3 picked · head to head
GPT-5.1 vs GPT-5.4 vs Grok 4.20
Side by side · benchmarks, pricing, and signals you can act on.
Winner summary
GPT-5.4 wins on 22/25 benchmarks
GPT-5.4 wins 22 of 25 shared benchmarks. Leads in reasoning · knowledge · general.
Category leads
reasoning·GPT-5.4knowledge·GPT-5.4general·GPT-5.4math·GPT-5.4coding·GPT-5.4agentic·GPT-5.4arena·GPT-5.4
Hype vs Reality
Attention vs performance
GPT-5.1
#138 by perf·#4 by attention
GPT-5.4
#104 by perf·#4 by attention
Grok 4.20
#129 by perf·#15 by attention
Best value
Grok 4.20
3.0x better value than GPT-5.1
GPT-5.1
8.6 pts/$
$5.63/M
GPT-5.4
6.0 pts/$
$8.75/M
Grok 4.20
26.0 pts/$
$1.88/M
Vendor risk
Who is behind the model
OpenAI
$840.0B·Tier 1
OpenAI
$840.0B·Tier 1
xAI
$250.0B·Tier 1
Head to head
25 benchmarks · 3 models
GPT-5.1GPT-5.4Grok 4.20
ARC-AGI
GPT-5.4 leads by +4.2
ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization.
GPT-5.1
72.8
GPT-5.4
93.7
Grok 4.20
89.5
ARC-AGI-2
GPT-5.4 leads by +8.8
ARC-AGI-2 · the second iteration of the Abstraction and Reasoning Corpus, testing novel pattern recognition and abstract reasoning without prior training data.
GPT-5.1
17.6
GPT-5.4
74.0
Grok 4.20
65.1
Chess Puzzles
GPT-5.4 leads by +12.6
Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities.
GPT-5.1
28.4
GPT-5.4
41.1
Grok 4.20
20.0
Cl Bench
GPT-5.4 leads by +4.2
GPT-5.1
23.7
GPT-5.4
27.9
Grok 4.20
22.2
Cl Bench Life
GPT-5.4 leads by +4.4
GPT-5.1
17.3
GPT-5.4
21.7
Grok 4.20
11.9
Dtbench
GPT-5.4 leads by +7.1
GPT-5.1
83.5
GPT-5.4
90.7
Grok 4.20
83.5
GPQA diamond
GPT-5.4 leads by +5.3
Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs.
GPT-5.1
83.5
GPT-5.4
91.1
Grok 4.20
85.8
Lmca
GPT-5.4 leads by +9.5
GPT-5.1
51.6
GPT-5.4
61.1
Grok 4.20
45.5
OTIS Mock AIME 2024-2025
GPT-5.4 leads by +5.6
OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills.
GPT-5.1
88.6
GPT-5.4
97.8
Grok 4.20
92.2
SimpleQA Verified
GPT-5.1 leads by +2.9
SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information.
GPT-5.1
48.0
GPT-5.4
45.1
Grok 4.20
30.2
Terminal Bench
GPT-5.4 leads by +24.5
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-5.1
47.6
GPT-5.4
81.8
Grok 4.20
57.3
WeirdML
GPT-5.4 leads by +16.9
WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns.
GPT-5.1
60.8
GPT-5.4
77.7
Grok 4.20
52.3
APEX-Agents
GPT-5.4 leads by +34.9
APEX-Agents · evaluates AI agents on complex, multi-step tasks requiring planning, tool use, and autonomous decision-making in realistic environments.
GPT-5.1
17.5
GPT-5.4
52.4
Chatbot Arena Elo · Coding
GPT-5.4 leads by +54.0
GPT-5.1
1341.0
GPT-5.4
1395.0
Chatbot Arena Elo · Overall
GPT-5.4 leads by +26.3
GPT-5.1
1438.5
GPT-5.4
1464.8
DeepResearch Bench
GPT-5.1 leads by +7.7
DeepResearch Bench · evaluates AI on complex multi-step research tasks requiring information gathering, synthesis, and producing comprehensive analyses.
GPT-5.1
42.8
GPT-5.4
35.1
FrontierMath-2025-02-28-Private
GPT-5.1 leads by +6.9
FrontierMath (Feb 2025) · original research-level math problems created by mathematicians, testing capabilities at the boundary of current AI mathematical reasoning.
GPT-5.1
54.5
GPT-5.4
47.6
FrontierMath-Tier-4-2025-07-01-Private
GPT-5.4 leads by +6.3
FrontierMath Tier 4 (Jul 2025) · the most challenging tier of frontier mathematics, containing problems that push the absolute limits of AI mathematical reasoning.
GPT-5.1
20.8
GPT-5.4
27.1
FrontierMath-Tier-4-v2-Private
GPT-5.4 leads by +31.9
GPT-5.4
49.0
Grok 4.20
17.1
FrontierMath-Tiers-1-3-v2-Private
GPT-5.4 leads by +33.7
GPT-5.4
78.6
Grok 4.20
44.9
GSO-Bench
GPT-5.4 leads by +17.6
GSO-Bench · evaluates AI models on real-world open-source software engineering tasks, testing the ability to understand and resolve actual GitHub issues.
GPT-5.1
13.7
GPT-5.4
31.4
HLE
GPT-5.4 leads by +13.2
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-5.1
19.8
GPT-5.4
33.0
Mystery Game Puzzles
GPT-5.4 leads by +19.8
GPT-5.1
10.8
GPT-5.4
30.6
Proofbench
GPT-5.4 leads by +42.0
GPT-5.4
56.0
Grok 4.20
14.0
SWE-Bench verified
GPT-5.4 leads by +8.9
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-5.1
68.0
GPT-5.4
76.9
Full benchmark table
| Benchmark | GPT-5.1 | GPT-5.4 | Grok 4.20 |
|---|---|---|---|
ARC-AGI ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization. | 72.8 | 93.7 | 89.5 |
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. | 17.6 | 74.0 | 65.1 |
Chess Puzzles Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities. | 28.4 | 41.1 | 20.0 |
Cl Bench | 23.7 | 27.9 | 22.2 |
Cl Bench Life | 17.3 | 21.7 | 11.9 |
Dtbench | 83.5 | 90.7 | 83.5 |
GPQA diamond Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs. | 83.5 | 91.1 | 85.8 |
Lmca | 51.6 | 61.1 | 45.5 |
OTIS Mock AIME 2024-2025 OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills. | 88.6 | 97.8 | 92.2 |
SimpleQA Verified SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information. | 48.0 | 45.1 | 30.2 |
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. | 47.6 | 81.8 | 57.3 |
WeirdML WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns. | 60.8 | 77.7 | 52.3 |
APEX-Agents APEX-Agents · evaluates AI agents on complex, multi-step tasks requiring planning, tool use, and autonomous decision-making in realistic environments. | 17.5 | 52.4 | — |
Chatbot Arena Elo · Coding | 1341.0 | 1395.0 | — |
Chatbot Arena Elo · Overall | 1438.5 | 1464.8 | — |
DeepResearch Bench DeepResearch Bench · evaluates AI on complex multi-step research tasks requiring information gathering, synthesis, and producing comprehensive analyses. | 42.8 | 35.1 | — |
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. | 54.5 | 47.6 | — |
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. | 20.8 | 27.1 | — |
FrontierMath-Tier-4-v2-Private | — | 49.0 | 17.1 |
FrontierMath-Tiers-1-3-v2-Private | — | 78.6 | 44.9 |
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. | 13.7 | 31.4 | — |
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%. | 19.8 | 33.0 | — |
Mystery Game Puzzles | 10.8 | 30.6 | — |
Proofbench | — | 56.0 | 14.0 |
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. 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. | 68.0 | 76.9 | — |
Pricing · per 1M tokens · projected $/mo at 10M tokens