vs
1:15
overall capability wins (0 ties across 16 scored). GLM-5.3 leads overall.
- Reasoning1:4
- Coding0:5
- Arena0:1
- Agents0:1
- Tool use0:2
- Composite indices0:2
- Specs0:2
Verdict
GLM-5.3 leads coding (widest gap: +23.7 on TermBench 3); GLM-5.3 ranks higher on LMArena (+11 Elo); GLM-5.3 offers ~5.0x larger context (1M vs 200K).
- GLM-5.3 leads coding (widest gap: +23.7 on TermBench 3)
- GLM-5.3 ranks higher on LMArena (+11 Elo)
- GLM-5.3 offers ~5.0x larger context (1M vs 200K)
- GLM-5.3 allows ~3.9x more max output (128K vs 33K)
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- GLM-5.2
- $2.50
- GLM-5.3
- -
Price per 1M tokens
Full comparison
Organization
Zhipu AI
Zhipu AI
Family
GLM
GLM
License
MIT
GLM-5.3 License
Open weights
Yes
Yes
Release
Jun 1, 2026
Aug 14, 2026
Knowledge cutoff
-
-
API / provider
Zhipu / Z.ai
No primary API price listed
Modalities
text, image → text
text → text
Specs
Context window
200K
1M
Max output
33K
128K
Parameters
—
744B (40B act.)
Pricing
Input $/1M
$1.40
—
Output $/1M
$4.40
—
Blended $/1M (3∶1)
$2.15
-
Speed
tok/s
168
-
TTFT (s)
0.35
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
89.5
91.7
Humanity's Last Exam
41.1
42.3
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
24.3
33.9
AA-LCR v1.1
78.3
79.7
CritPt
20.9
19.1
MMMU-Pro
—
—
IFBench
73.3
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
—
Terminal-Bench 2.1
77.9
83.9
Aider Polyglot
—
—
Terminal-Bench 3
4.6
28.3
BigCodeBench
—
—
SciCode
51.2
59.0
CursorBench
55.0
—
SWE-Rebench
57.0
—
NL2Repo-Bench
—
58.0
DeepSWE
44.0
66.9
WebDev Arena
1,592
1,614
Terminal-Bench 4.0
—
41.8
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,472
1,483
| Metric | GLM-5.2 | GLM-5.3 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Zhipu AI | Zhipu AI | — |
| Family | GLM | GLM | — |
| License | MIT | GLM-5.3 License | — |
| Open weights | Yes | Yes | — |
| Release | Jun 1, 2026 | Aug 14, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Zhipu / Z.ai | No primary API price listed | — |
| Modalities | text, image → text | text → text | — |
| Specs | |||
| Context window | 200K | 1M | B +800K |
| Max output | 33K | 128K | B +95K |
| Parameters | — | 744B (40B act.) | — |
| Pricing | |||
| Input $/1M | $1.40 | — | — |
| Output $/1M | $4.40 | — | — |
| Blended $/1M (3∶1) | $2.15 | - | — |
| Speed | |||
| tok/s | 168 | - | — |
| TTFT (s) | 0.35 | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | 89.5 | 91.7 | B +2.2 pts |
| Humanity's Last Exam | 41.1 | 42.3 | B +1.2 pts |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 24.3 | 33.9 | B +9.6 pts |
| AA-LCR v1.1 | 78.3 | 79.7 | B +1.4 pts |
| CritPt | 20.9 | 19.1 | A +1.8 pts |
| MMMU-Pro | — | — | — |
| IFBench | 73.3 | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | 77.9 | 83.9 | B +6.0 pts |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | 4.6 | 28.3 | B +23.7 pts |
| BigCodeBench | — | — | — |
| SciCode | 51.2 | 59.0 | B +7.8 pts |
| CursorBench | 55.0 | — | — |
| SWE-Rebench | 57.0 | — | — |
| NL2Repo-Bench | — | 58.0 | — |
| DeepSWE | 44.0 | 66.9 | B +22.9 pts |
| WebDev Arena | 1,592 | 1,614 | B +23 Elo |
| Terminal-Bench 4.0 | — | 41.8 | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,472 | 1,483 | B +11 Elo |
GLM-5.2
- Context
- 200K / 33K out
- Parameters
- —
- Price
- $1.40 / $4.40
- Speed
- 168 tok/s · 0.35s TTFT
- Modalities
- text, image → text
- License
- MIT
Open MIT GLM-5.2 - competitive open Arena Elo and strong Chinese/English bilingual performance.
GLM-5.3
- Context
- 1M / 128K out
- Parameters
- 744B (40B act.)
- Price
- —
- Speed
- —
- Modalities
- text → text
- License
- GLM-5.3 License
Zhipu's Aug 14, 2026 flagship — GLM-5.2 base re-post-trained for complex software engineering, terminal, and real-world agent tasks, with SOTA open-weights coding and emergent cybersecurity skills (CyberGym 84.5). Open weights (744B MoE / 40B active) are on Hugging Face under the GLM-5.3 license; thinking effort is low/high/max (default max).
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.