Compare · ModelsLive · 3 picked · head to head
DeepSeek V4 Pro vs GLM 5.2 vs GPT-5.2-Codex
Side by side · benchmarks, pricing, and signals you can act on.
Winner summary
GLM 5.2 wins on 19/27 benchmarks
GLM 5.2 wins 19 of 27 shared benchmarks. Leads in coding · knowledge · speed.
Category leads
coding·GLM 5.2reasoning·GPT-5.2-Codexlanguage·GPT-5.2-Codexmath·DeepSeek V4 Proknowledge·GLM 5.2speed·GLM 5.2arena·GLM 5.2general·GLM 5.2
Hype vs Reality
Attention vs performance
DeepSeek V4 Pro
#89 by perf·#6 by attention
GLM 5.2
#67 by perf·#3 by attention
GPT-5.2-Codex
#17 by perf·#4 by attention
Best value
DeepSeek V4 Pro
7.6x better value than GLM 5.2
DeepSeek V4 Pro
172.1 pts/$
$0.31/M
GLM 5.2
22.6 pts/$
$2.50/M
GPT-5.2-Codex
9.0 pts/$
$7.88/M
Vendor risk
Mixed exposure
One or more vendors flagged
DeepSeek
$3.4B·Tier 1
z-ai
private · undisclosed
OpenAI
$840.0B·Tier 1
Head to head
27 benchmarks · 3 models
DeepSeek V4 ProGLM 5.2GPT-5.2-Codex
LiveBench · Agentic Coding
GLM 5.2 leads by +16.7
DeepSeek V4 Pro
56.7
GLM 5.2
73.3
GPT-5.2-Codex
51.7
LiveBench · Coding
GPT-5.2-Codex leads by +4.0
DeepSeek V4 Pro
70.0
GLM 5.2
79.7
GPT-5.2-Codex
83.6
LiveBench · Data Analysis
GPT-5.2-Codex leads by +3.7
DeepSeek V4 Pro
74.5
GLM 5.2
73.7
GPT-5.2-Codex
78.2
LiveBench · If
GPT-5.2-Codex leads by +4.1
DeepSeek V4 Pro
62.4
GLM 5.2
62.3
GPT-5.2-Codex
66.5
LiveBench · Language
DeepSeek V4 Pro leads by +1.9
DeepSeek V4 Pro
78.1
GLM 5.2
76.2
GPT-5.2-Codex
73.7
LiveBench · Mathematics
DeepSeek V4 Pro leads by +0.9
DeepSeek V4 Pro
90.7
GLM 5.2
89.8
GPT-5.2-Codex
88.8
LiveBench · Overall
GLM 5.2 leads by +1.9
DeepSeek V4 Pro
73.6
GLM 5.2
76.2
GPT-5.2-Codex
74.3
LiveBench · Reasoning
DeepSeek V4 Pro leads by +4.1
DeepSeek V4 Pro
82.7
GLM 5.2
78.6
GPT-5.2-Codex
77.7
Artificial Analysis · Agentic Index
GLM 5.2 leads by +6.7
Artificial Analysis Agentic Index · a composite score measuring how well a model performs in agentic workflows · multi-step tool use, planning, error recovery, and autonomous task completion. Aggregates results from multiple agentic benchmarks including SWE-bench, tool-use tests, and planning evaluations. The canonical single-number metric for "how good is this model as an agent?"
DeepSeek V4 Pro
36.4
GLM 5.2
43.1
Artificial Analysis · Coding Index
GLM 5.2 leads by +9.4
Artificial Analysis Coding Index · a composite score that aggregates performance across multiple coding benchmarks into a single index. Tracks code generation quality, debugging ability, multi-language competence, and real-world software engineering tasks. Used by Artificial Analysis to rank model coding capability in a normalized, comparable format. Useful for developers choosing between models for coding-heavy workloads.
DeepSeek V4 Pro
59.4
GLM 5.2
68.8
Artificial Analysis · Quality Index
GLM 5.2 leads by +15.1
DeepSeek V4 Pro
36.0
GLM 5.2
51.1
Chatbot Arena Elo · Coding
GLM 5.2 leads by +147.3
DeepSeek V4 Pro
1446.0
GLM 5.2
1593.3
Chatbot Arena Elo · Overall
GLM 5.2 leads by +13.2
DeepSeek V4 Pro
1457.8
GLM 5.2
1471.0
Chess Puzzles
GLM 5.2 leads by +1.1
Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities.
DeepSeek V4 Pro
15.8
GLM 5.2
16.9
Dtbench
GLM 5.2 leads by +4.9
DeepSeek V4 Pro
84.5
GLM 5.2
89.3
Frontiercode
GLM 5.2 leads by +6.9
DeepSeek V4 Pro
17.6
GLM 5.2
24.5
FrontierMath-Tier-4-v2-Private
GLM 5.2 leads by +26.8
DeepSeek V4 Pro
2.4
GLM 5.2
29.3
FrontierMath-Tiers-1-3-v2-Private
GLM 5.2 leads by +14.0
DeepSeek V4 Pro
45.3
GLM 5.2
59.2
GPQA diamond
GLM 5.2 leads by +1.3
Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs.
DeepSeek V4 Pro
87.9
GLM 5.2
89.1
Lmca
GLM 5.2 leads by +5.4
DeepSeek V4 Pro
48.5
GLM 5.2
53.9
Mystery Game Puzzles
GLM 5.2 leads by +2.2
DeepSeek V4 Pro
8.6
GLM 5.2
10.8
OTIS Mock AIME 2024-2025
DeepSeek V4 Pro leads by +10.3
OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills.
DeepSeek V4 Pro
96.7
GLM 5.2
86.4
Proofbench
GLM 5.2 leads by +19.0
DeepSeek V4 Pro
16.0
GLM 5.2
35.0
SimpleQA Verified
DeepSeek V4 Pro leads by +12.8
SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information.
DeepSeek V4 Pro
47.0
GLM 5.2
34.2
Surface Evolver Bench
GLM 5.2 leads by +15.6
DeepSeek V4 Pro
40.0
GLM 5.2
55.6
SWE-Bench verified
GLM 5.2 leads by +1.1
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.
DeepSeek V4 Pro
77.6
GLM 5.2
78.7
WeirdML
GLM 5.2 leads by +21.2
WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns.
DeepSeek V4 Pro
48.9
GLM 5.2
70.1
Full benchmark table
| Benchmark | DeepSeek V4 Pro | GLM 5.2 | GPT-5.2-Codex |
|---|---|---|---|
LiveBench · Agentic Coding | 56.7 | 73.3 | 51.7 |
LiveBench · Coding | 70.0 | 79.7 | 83.6 |
LiveBench · Data Analysis | 74.5 | 73.7 | 78.2 |
LiveBench · If | 62.4 | 62.3 | 66.5 |
LiveBench · Language | 78.1 | 76.2 | 73.7 |
LiveBench · Mathematics | 90.7 | 89.8 | 88.8 |
LiveBench · Overall | 73.6 | 76.2 | 74.3 |
LiveBench · Reasoning | 82.7 | 78.6 | 77.7 |
Artificial Analysis · Agentic Index Artificial Analysis Agentic Index · a composite score measuring how well a model performs in agentic workflows · multi-step tool use, planning, error recovery, and autonomous task completion. Aggregates results from multiple agentic benchmarks including SWE-bench, tool-use tests, and planning evaluations. The canonical single-number metric for "how good is this model as an agent?" | 36.4 | 43.1 | — |
Artificial Analysis · Coding Index Artificial Analysis Coding Index · a composite score that aggregates performance across multiple coding benchmarks into a single index. Tracks code generation quality, debugging ability, multi-language competence, and real-world software engineering tasks. Used by Artificial Analysis to rank model coding capability in a normalized, comparable format. Useful for developers choosing between models for coding-heavy workloads. | 59.4 | 68.8 | — |
Artificial Analysis · Quality Index | 36.0 | 51.1 | — |
Chatbot Arena Elo · Coding | 1446.0 | 1593.3 | — |
Chatbot Arena Elo · Overall | 1457.8 | 1471.0 | — |
Chess Puzzles Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities. | 15.8 | 16.9 | — |
Dtbench | 84.5 | 89.3 | — |
Frontiercode | 17.6 | 24.5 | — |
FrontierMath-Tier-4-v2-Private | 2.4 | 29.3 | — |
FrontierMath-Tiers-1-3-v2-Private | 45.3 | 59.2 | — |
GPQA diamond Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs. | 87.9 | 89.1 | — |
Lmca | 48.5 | 53.9 | — |
Mystery Game Puzzles | 8.6 | 10.8 | — |
OTIS Mock AIME 2024-2025 OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills. | 96.7 | 86.4 | — |
Proofbench | 16.0 | 35.0 | — |
SimpleQA Verified SimpleQA Verified · short factual questions with verified answers, measuring factual accuracy and the tendency to hallucinate or provide incorrect information. | 47.0 | 34.2 | — |
Surface Evolver Bench | 40.0 | 55.6 | — |
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. | 77.6 | 78.7 | — |
WeirdML WeirdML · tests models on unusual and adversarial machine learning tasks that require creative problem-solving beyond standard patterns. | 48.9 | 70.1 | — |
Pricing · per 1M tokens · projected $/mo at 10M tokens
| Model | Input | Output | Context | Projected $/mo |
|---|---|---|---|---|
| $0.21 | $0.42 | 1.0M tokens (~524 books) | $2.61 | |
| $1.00 | $4.00 | 1.0M tokens (~524 books) | $17.50 | |
| $1.75 | $14.00 | 400K tokens (~200 books) | $48.13 |