Compare · ModelsLive · 3 picked · head to head
GLM 4.7 vs GLM 5 vs Step 3.5 Flash
Side by side · benchmarks, pricing, and signals you can act on.
Winner summary
GLM 5 wins on 23/26 benchmarks
GLM 5 wins 23 of 26 shared benchmarks. Leads in arena · math · knowledge.
Category leads
arena·GLM 5math·GLM 5knowledge·GLM 5language·GLM 5coding·GLM 5agentic·GLM 5general·GLM 5reasoning·GLM 5
Hype vs Reality
Attention vs performance
GLM 4.7
#152 by perf·#3 by attention
GLM 5
#79 by perf·#3 by attention
Step 3.5 Flash
#8 by perf·no signal
Best value
Step 3.5 Flash
8.8x better value than GLM 5
GLM 4.7
32.9 pts/$
$1.40/M
GLM 5
43.7 pts/$
$1.26/M
Step 3.5 Flash
384.5 pts/$
$0.20/M
Vendor risk
Mixed exposure
One or more vendors flagged
z-ai
private · undisclosed
z-ai
private · undisclosed
StepFun
$5.0B·Tier 1
Head to head
26 benchmarks · 3 models
GLM 4.7GLM 5Step 3.5 Flash
Chatbot Arena Elo · Overall
GLM 5 leads by +16.1
GLM 4.7
1441.5
GLM 5
1457.6
Step 3.5 Flash
1393.3
OpenCompass · AIME2025
GLM 5 leads by +0.1
GLM 4.7
95.4
GLM 5
95.8
Step 3.5 Flash
95.7
OpenCompass · GPQA-Diamond
GLM 4.7 leads by +1.6
GLM 4.7
86.9
GLM 5
85.3
Step 3.5 Flash
83.7
OpenCompass · HLE
GLM 5 leads by +2.7
GLM 4.7
25.4
GLM 5
28.1
Step 3.5 Flash
21.6
OpenCompass · IFEval
GLM 4.7
90.2
GLM 5
93.2
Step 3.5 Flash
93.2
OpenCompass · LiveCodeBenchV6
GLM 5 leads by +2.3
GLM 4.7
83.8
GLM 5
86.2
Step 3.5 Flash
83.9
OpenCompass · MMLU-Pro
GLM 5 leads by +1.2
GLM 4.7
84.0
GLM 5
85.2
Step 3.5 Flash
83.5
APEX-Agents
GLM 5 leads by +14.1
APEX-Agents · evaluates AI agents on complex, multi-step tasks requiring planning, tool use, and autonomous decision-making in realistic environments.
GLM 4.7
3.1
GLM 5
17.2
Chatbot Arena Elo · Coding
GLM 4.7 leads by +0.5
GLM 4.7
1434.5
GLM 5
1434.0
Chess Puzzles
GLM 5 leads by +4.2
Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities.
GLM 4.7
1.1
GLM 5
5.3
Cl Bench
GLM 5 leads by +2.8
GLM 4.7
15.9
GLM 5
18.7
FrontierMath-2025-02-28-Private
GLM 5 leads by +24.5
FrontierMath (Feb 2025) · original research-level math problems created by mathematicians, testing capabilities at the boundary of current AI mathematical reasoning.
GLM 4.7
4.3
GLM 5
28.8
FrontierMath-Tier-4-2025-07-01-Private
GLM 5 leads by +3.5
FrontierMath Tier 4 (Jul 2025) · the most challenging tier of frontier mathematics, containing problems that push the absolute limits of AI mathematical reasoning.
GLM 4.7
0.0
GLM 5
3.5
GPQA diamond
GLM 5 leads by +6.0
Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs.
GLM 4.7
77.8
GLM 5
83.8
LiveBench · Agentic Coding
GLM 5 leads by +13.3
GLM 4.7
41.7
GLM 5
55.0
LiveBench · Coding
GLM 5 leads by +0.5
GLM 4.7
73.1
GLM 5
73.6
LiveBench · Data Analysis
GLM 5 leads by +12.7
GLM 4.7
55.2
GLM 5
67.9
LiveBench · If
GLM 5 leads by +19.7
GLM 4.7
35.7
GLM 5
55.3
LiveBench · Language
GLM 5 leads by +12.3
GLM 4.7
65.2
GLM 5
77.5
LiveBench · Mathematics
GLM 5 leads by +7.4
GLM 4.7
76.0
GLM 5
83.5
LiveBench · Overall
GLM 5 leads by +10.8
GLM 4.7
58.1
GLM 5
68.8
LiveBench · Reasoning
GLM 5 leads by +9.4
GLM 4.7
59.7
GLM 5
69.1
OTIS Mock AIME 2024-2025
GLM 4.7 leads by +3.3
OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills.
GLM 4.7
83.3
GLM 5
80.0
PostTrainBench
GLM 5 leads by +6.4
GLM 4.7
7.5
GLM 5
13.9
SimpleBench
GLM 5 leads by +6.6
SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking.
GLM 4.7
37.2
GLM 5
43.8
Terminal Bench
GLM 5 leads by +19.0
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.
GLM 4.7
33.4
GLM 5
52.4
Full benchmark table
| Benchmark | GLM 4.7 | GLM 5 | Step 3.5 Flash |
|---|---|---|---|
Chatbot Arena Elo · Overall | 1441.5 | 1457.6 | 1393.3 |
OpenCompass · AIME2025 | 95.4 | 95.8 | 95.7 |
OpenCompass · GPQA-Diamond | 86.9 | 85.3 | 83.7 |
OpenCompass · HLE | 25.4 | 28.1 | 21.6 |
OpenCompass · IFEval | 90.2 | 93.2 | 93.2 |
OpenCompass · LiveCodeBenchV6 | 83.8 | 86.2 | 83.9 |
OpenCompass · MMLU-Pro | 84.0 | 85.2 | 83.5 |
APEX-Agents APEX-Agents · evaluates AI agents on complex, multi-step tasks requiring planning, tool use, and autonomous decision-making in realistic environments. | 3.1 | 17.2 | — |
Chatbot Arena Elo · Coding | 1434.5 | 1434.0 | — |
Chess Puzzles Chess Puzzles · tests strategic and tactical reasoning by having models solve chess puzzle positions, evaluating lookahead and pattern recognition abilities. | 1.1 | 5.3 | — |
Cl Bench | 15.9 | 18.7 | — |
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. | 4.3 | 28.8 | — |
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. | 0.0 | 3.5 | — |
GPQA diamond Graduate-Level Google-Proof QA (Diamond set) · expert-crafted questions in physics, biology, and chemistry that are difficult even for domain PhDs. | 77.8 | 83.8 | — |
LiveBench · Agentic Coding | 41.7 | 55.0 | — |
LiveBench · Coding | 73.1 | 73.6 | — |
LiveBench · Data Analysis | 55.2 | 67.9 | — |
LiveBench · If | 35.7 | 55.3 | — |
LiveBench · Language | 65.2 | 77.5 | — |
LiveBench · Mathematics | 76.0 | 83.5 | — |
LiveBench · Overall | 58.1 | 68.8 | — |
LiveBench · Reasoning | 59.7 | 69.1 | — |
OTIS Mock AIME 2024-2025 OTIS Mock AIME 2024-2025 · simulated American Invitational Mathematics Examination problems testing advanced problem-solving skills. | 83.3 | 80.0 | — |
PostTrainBench | 7.5 | 13.9 | — |
SimpleBench SimpleBench · tests fundamental reasoning capabilities with straightforward problems designed to expose gaps in basic logical and spatial thinking. | 37.2 | 43.8 | — |
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. | 33.4 | 52.4 | — |
Pricing · per 1M tokens · projected $/mo at 10M tokens
| Model | Input | Output | Context | Projected $/mo |
|---|---|---|---|---|
| $0.60 | $2.20 | 205K tokens (~102 books) | $10.00 | |
| $0.60 | $1.92 | 205K tokens (~102 books) | $9.30 | |
| $0.10 | $0.30 | 262K tokens (~131 books) | $1.50 |