GLM 5.2 vs GLM 5.3
Different efforts: non-thinking · low · Compare all at max
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
GLM 5.2 · max28.4
GLM 5.352
GDPval-AA (work tasks) 2.1 GDPval-AA (work tasks) 2.1: Real work tasks from many professions (reports, spreadsheets, analyses), graded against expert work
GLM 5.237.4
GLM 5.341.1
τ³-Banking · AA τ³-Banking · AA: Agents retrieve banking knowledge and execute tools to resolve customer-support workflows
GLM 5.216.7
GLM 5.3 · max50.3
ITBench SRE ITBench SRE: IT incident tasks: diagnose and fix broken systems like a site-reliability engineer
GLM 5.2 · max42.7
GLM 5.3 · max46.1
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
GLM 5.2 · max1
GLM 5.334.8
AA-Briefcase 1.1 · Elo AA-Briefcase 1.1 · Elo: Agents produce professional deliverables; blind pairwise comparisons produce the source Elo rating
GLM 5.2 · max1,228.9
GLM 5.31,186.8
AA-LCR (long context) 1.1 AA-LCR (long context) 1.1: Questions that need reading and connecting facts across very long documents (around 100K tokens)
GLM 5.242.3
GLM 5.372.7
DeepSWE 1.1 DeepSWE 1.1: Agents change repository code to resolve software engineering tasks under the stated DeepSWE edition
GLM 5.2 · max43.8
GLM 5.3 · max69
EnterpriseOps-Gym-AA EnterpriseOps-Gym-AA: Agents complete stateful enterprise workflows in eight domains; final database state determines success
GLM 5.2 · max42.7
GLM 5.3 · max36.4
τ³-bench · banking · 1.0.1 τ-bench official · 3-banking-1.0.1: Customer-service agents converse with simulated users, retrieve domain knowledge and change state through tools
GLM 5.2 · xhigh37.1
GLM 5.3—
AA Analyst Agent AA Analyst Agent: Data-analyst tasks: the model explores data with tools and reports findings
GLM 5.2—
GLM 5.3 · max38.8
APEX-Agents APEX-Agents: Agents complete long-horizon professional-services tasks across realistic applications
GLM 5.2 · max33.7
GLM 5.3—
Terminal-Bench Hard (version unstated) Terminal-Bench Hard (version unstated): Agents solve difficult software, system-administration and data-processing tasks in a terminal
GLM 5.2 · max50.8
GLM 5.3—
Terminal-Bench Science 0.1 Terminal-Bench Science 0.1: Agents complete expert-authored research workflows across five scientific domains in a terminal
GLM 5.2—
GLM 5.3 · max9.5
FrontierCode 1.1 FrontierCode 1.1: Agents implement changes in real repositories, graded against maintainer-defined requirements
GLM 5.2—
GLM 5.3 · max40.1
Harvey LAB-AA 1.1 Harvey LAB-AA 1.1: Legal agents work from case documents; every rubric criterion must pass and material hallucinations receive no credit
GLM 5.2—
GLM 5.3 · max0.3
SWE-bench Pro (v2 full) SWE-bench Pro (v2 full): Agents resolve repository issues under the network-locked v2 protocol; full and 51-task Hard results are separate
GLM 5.2—
GLM 5.3 · max95.6
SWE-bench Pro (v2 hard) SWE-bench Pro (v2 hard): Agents resolve repository issues under the network-locked v2 protocol; full and 51-task Hard results are separate
GLM 5.2—
GLM 5.3 · max84.3
EBR Bench EBR Bench: Agents learn to play Earthborne Rangers across repeated playthroughs
GLM 5.2 · max9.5
GLM 5.3—
| Benchmark | GLM 5.2 | GLM 5.3 |
|---|---|---|
| AutomationBench-AA | 28.4 · max | 52 |
| GDPval-AA (work tasks) 2.1 | 37.4 | 41.1 |
| τ³-Banking · AA | 16.7 | 50.3 · max |
| ITBench SRE | 42.7 · max | 46.1 · max |
| Terminal-Bench 4.0 | 1 · max | 34.8 |
| AA-Briefcase 1.1 · Elo | 1,228.9 · max | 1,186.8 |
| AA-LCR (long context) 1.1 | 42.3 | 72.7 |
| DeepSWE 1.1 | 43.8 · max | 69 · max |
| EnterpriseOps-Gym-AA | 42.7 · max | 36.4 · max |
| τ³-bench · banking · 1.0.1 | 37.1 · xhigh | — |
| AA Analyst Agent | — | 38.8 · max |
| APEX-Agents | 33.7 · max | — |
| Terminal-Bench Hard (version unstated) | 50.8 · max | — |
| Terminal-Bench Science 0.1 | — | 9.5 · max |
| FrontierCode 1.1 | — | 40.1 · max |
| Harvey LAB-AA 1.1 | — | 0.3 · max |
| SWE-bench Pro (v2 full) | — | 95.6 · max |
| SWE-bench Pro (v2 hard) | — | 84.3 · max |
| EBR Bench | 9.5 · max | — |
Specs
| Spec | GLM 5.2 | GLM 5.3 |
|---|---|---|
| Input, per 1M | !$1.40 | !$1.40 |
| Output, per 1M | !$4.40 | !$4.40 |
| Cost per task? | — | $0.85 |
| Speed, tok/s | 171 | 126 |
| Context | 1M | 1M |
| Weights | Open | Open |
| Released | Jun 2026 | Aug 2026 |
| Memory (4-bit) | 452 GB | 452 GB |
| Licence | MIT | glm-5.3 |
| Commercial use | Allowed | Restricted |
| See images | — | — |
| Read PDFs | — | — |
| Hear audio | — | — |
| Watch video | — | — |
| Call tools | ✓ | ✓ |
| Force a tool call | Partial | — |
| Return strict JSON | ✓ | ✓ |
| Search the web | ? | ? |
| Run code | ? | ? |
| Use a computer | ? | ? |
What breaks when you switch
GLM 5.2 → GLM 5.3
Reasoning cannot be disabled›
Disabling reasoning is unsupported; keep reasoning enabled.
tool_choice auto · Z.ai API›
The first-party chat reference exposes only auto for tool_choice.
Complete unmodified reasoning_content›
Preserved thinking defaults on for Coding Plan and off for the standard API; when enabled, return complete unmodified reasoning_content in its original order to protect quality and cache hits.
Only one stop word›
Only one stop word is supported, so do not rely on a list of alternative stop sequences.
GLM 5.3 → GLM 5.2
Nothing known.