Inference
EN

GLM-5.3 vs Hy4 preview

Change models
GLM-5.3Z.AI
Hy4 previewTencent
IndexAll categories, out of 100.
65.8 ±3.5
66.3 ±4.1 best
GapIndex points below the top.
−0.5 · no clear gap
Top
CoverageIndex weight that has results.
85%
90% best
SpeedOutput tokens per second.
86/s best
67/s
EloLMArena rating and rank.
1476
—
ReleasedMonth of release.
2026-08
2026-08
Input / 1MUS dollars per 1M input tokens.
$0.07 best
$0.834
Output / 1MUS dollars per 1M output tokens.
$1.40 best
$2.50
Batch output / 1MBatch price per 1M output tokens.
$0.989
—
1K chats2K tokens in, 500 out, each.
$0.84 best
$2.92
1K documents20K tokens in, 1K out, each.
$2.80 best
$19.18
1K agent tasks200K tokens in, 20K out, each.
$41.98 best
$216.82
ContextMaximum tokens in one request.
1.05M
1.05M
AgenticMulti-step tasks with tools.
68.1 best
63.4
CodingCode writing and repair.
66.0 best
65.4
ReasoningLogic problems and puzzles.
64.5
69.4 best
MultimodalTasks with images and text.
—
75.5
KnowledgeFacts and expert knowledge.
59.1
67.7 best
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
66.7
—
MathMath problems.
62.9
72.9 best

Benchmarks

37 shared
Only GLM-5.3: 60Only Hy4 preview: 16
GLM-5.3
Hy4 preview
Critical Physics TasksReasoning
19.1% best32 / 214
16.9%43 / 214
LegalBenchLegal
84.8% best29 / 149
83.8%47 / 149
Vibe Code Bench v1.1Coding
78.1% best36 / 123
77.5%37 / 123
MedScribeHealthcare
88.8% best11 / 109
83.6%45 / 109
MedCodeHealthcare
42.9%49 / 106
43.2% best47 / 106
Legal Research BenchLegal
49.0% best10 / 90
45.2%22 / 90
Terminal-Bench 2.1 · BenchLMAgentic
88.2% best10 / 86
85.4%18 / 86
Tax Agent BenchFinance
43.3% best7 / 82
32.6%21 / 82
Finance Agent (v2)Finance
55.8% best22 / 79
55.1%23 / 79
Code MigrationCoding
44.2%24 / 77
47.4% best14 / 77
EMBFinance
56.3%46 / 75
58.1% best38 / 75
Terminal-Bench 2.1 · Vals AIAgentic
71.5% best26 / 75
55.1%49 / 75
ProgramBenchCoding
1.5% best14 / 65
0.0%34 / 65
Terminal-Bench 4.0Agentic
38.9% best17 / 61
8.1%48 / 61
DeepSWEAgentic
66.9% best31 / 59
64.3%36 / 59
Agent Arena steerabilityAgentic
−1.232 / 54
0.2 best26 / 54
Agent Arena task outcomeAgentic
3.124 / 54
3.5 best21 / 54
ProofBench v1.1Math
49.0%34 / 54
75.0% best19 / 54
Agent Arena command recoveryAgentic
−0.134 / 50
6.5 best12 / 50
Public Benefits Bench v1.1Public benefits
68.5%10 / 50
68.6% best9 / 50