Compare models

Context, limits, capabilities, price and use cases side by side — up to 4 models at a time. The URL holds your selection, so you can share it.

GLM-5.3GLM-5.2 Add a modelChat with these 2

AttributeHide identical rows	GLM-5.3z-ai/glm-5.3	GLM-5.2z-ai/glm-5.2
Specs
Context window	1M	1M
Max output	128K	128K
Input modalities	Text	Text
Output modalities	Text	Text
Capabilities
Streaming	Supported	Supported
Tool calling	Supported	Supported
Structured output (JSON)	Supported	Supported
Image input	Not supported	Not supported
Thinking mode	Supported	Supported
Price / 1M tokens
Input	$ 1.20	$ 1.20
Output	$ 4.40	$ 4.40
Cache read	$ 0.30	$ 0.30
Cache write	$ 1.20	$ 1.20
What it is good at
Best for	Long-horizon software engineering Security-focused code review	Million-token workflows Agentic software engineering
Published benchmarks
Terminal-Bench 2.1pass_rate	—	81%Terminus-2
Terminal-Bench 3.0pass_rate	28.3%effort max · Claude Code 2.1.207 · 3 runs(setup not fully disclosed)	—
Agents' Last Exam (CLI)score	28.5%effort max(setup not fully disclosed)	—
DeepSWE 1.1resolved_rate	66.9%mini-swe-agent(setup not fully disclosed)	—
SWE-Bench Proresolved_rate	—	62.1%(setup not fully disclosed)
CyberGymscore	84.5%effort max · Claude Code 2.1.207(setup not fully disclosed)	—
ExploitBenchscore	54.4%effort max · Claude Code 2.1.207 · 3 runs(setup not fully disclosed)	—
ExploitGym (6h)tasks_completed	130effort max · Claude Code 2.1.207(setup not fully disclosed)	—
GDPval-AA v2elo	1769(setup not fully disclosed)	—
GPQA Diamondaccuracy	—	91.2%(setup not fully disclosed)
Humanity's Last Examaccuracy	—	40.5%
Humanity's Last Exam (with tools)accuracysetups differ by: reasoning, tools	62.5%effort max(setup not fully disclosed)	54.7%
MCP Atlas Publicscore	—	76.8%(setup not fully disclosed)
Tool Decathlonscore	—	48.2%(setup not fully disclosed)