Inference
EN

Hy4 vs Hy4 preview

Change models
Hy4Tencent
Hy4 previewTencent
IndexAll categories, out of 100.
72.0 ±7.8 best
66.4 ±4.1
GapIndex points below the top.
Top
−5.6 · no clear gap
CoverageIndex weight that has results.
65%
90% best
SpeedOutput tokens per second.
52/s
67/s best
EloLMArena rating and rank.
—
—
ReleasedMonth of release.
2026-11
2026-08
Input / 1MUS dollars per 1M input tokens.
$0.417 best
$0.834
Output / 1MUS dollars per 1M output tokens.
$1.25 best
$2.50
1K chats2K tokens in, 500 out, each.
$1.46 best
$2.92
1K documents20K tokens in, 1K out, each.
$9.59 best
$19.18
1K agent tasks200K tokens in, 20K out, each.
$108.41 best
$216.82
ContextMaximum tokens in one request.
1.05M
1.05M
AgenticMulti-step tasks with tools.
74.5 best
63.5
CodingCode writing and repair.
68.4 best
65.5
ReasoningLogic problems and puzzles.
—
69.6
MultimodalTasks with images and text.
57.4
75.7 best
KnowledgeFacts and expert knowledge.
83.7 best
67.9
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
—
—
MathMath problems.
—
73.0

Benchmarks

15 shared
Only Hy4: 6Only Hy4 preview: 38
Hy4
Hy4 preview
Vibe Code Bench v1.1Coding
81.0% best30 / 123
77.5%37 / 123
Legal Research BenchLegal
41.3%30 / 90
45.2% best22 / 90
Tax Agent BenchFinance
34.4% best17 / 82
32.6%21 / 82
Humanity's Last ExamKnowledge
59.1% best7 / 77
55.4%14 / 77
Terminal-Bench 4.0Agentic
39.4% best15 / 61
8.1%48 / 61
DeepSWEAgentic
74.9% best8 / 59
64.3%36 / 59
Agent Arena steerabilityAgentic
8.8 best14 / 54
0.226 / 54
Agent Arena task outcomeAgentic
6.7 best12 / 54
3.521 / 54
Humanity's Last Exam with toolsAgentic
59.1% best10 / 36
55.4%18 / 36
HLAB-Vals-1.1Legal
12.5% best17 / 35
8.8%23 / 35
AutomationBenchAgentic
56.2% best4 / 32
32.1%24 / 32
Agents' Last ExamAgentic
41.0% best8 / 21
22.8%21 / 21
ProgramBench: Can Language Models Rebuild Programs From Scratch?Coding
30.0% best12 / 19
17.5%19 / 19
APEX-AgentsAgentic
75.2% best1 / 13
37.1%8 / 13
SWE-MarathonCoding
43.8%
31.9%