Inference
EN

Hy4 vs Step 5 Preview

Change models
Hy4Tencent
Step 5 PreviewStepFun
IndexAll categories, out of 100.
72.0 ±7.8 best
68.4 ±3.7
GapIndex points below the top.
Top
−3.6 · no clear gap
CoverageIndex weight that has results.
65%
90% best
SpeedOutput tokens per second.
52/s
105/s best
EloLMArena rating and rank.
—
1443
ReleasedMonth of release.
2026-11
2026-09
Input / 1MUS dollars per 1M input tokens.
$0.417 best
$1.00
Output / 1MUS dollars per 1M output tokens.
$1.25 best
$2.70
1K chats2K tokens in, 500 out, each.
$1.46 best
$3.35
1K documents20K tokens in, 1K out, each.
$9.59 best
$22.70
1K agent tasks200K tokens in, 20K out, each.
$108.41 best
$254.00
ContextMaximum tokens in one request.
1.05M best
1M
AgenticMulti-step tasks with tools.
74.5 best
68.8
CodingCode writing and repair.
68.4
71.4 best
ReasoningLogic problems and puzzles.
—
67.4
MultimodalTasks with images and text.
57.4
62.2 best
KnowledgeFacts and expert knowledge.
83.7 best
63.5
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
—
—
MathMath problems.
—
61.7

Benchmarks

15 shared
Only Hy4: 6Only Step 5 Preview: 47
Hy4
Step 5 Preview
Vibe Code Bench v1.1Coding
81.0% best30 / 123
70.4%48 / 123
Legal Research BenchLegal
41.3% best30 / 90
40.4%35 / 90
Tax Agent BenchFinance
34.4% best17 / 82
27.3%39 / 82
Humanity's Last ExamKnowledge
59.1% best7 / 77
46.5%26 / 77
Terminal-Bench 4.0Agentic
39.4% best15 / 61
31.8%23 / 61
DeepSWEAgentic
74.9% best8 / 59
67.7%27 / 59
Agent Arena steerabilityAgentic
8.8 best14 / 54
−0.930 / 54
Agent Arena task outcomeAgentic
6.7 best12 / 54
−1.535 / 54
Massive Multi-discipline Multimodal Understanding ProMultimodal
78.6% best22 / 53
76.0%33 / 53
HLAB-Vals-1.1Legal
12.5%17 / 35
12.9% best15 / 35
AutomationBenchAgentic
56.2% best4 / 32
44.0%17 / 32
Agents' Last ExamAgentic
41.0% best8 / 21
29.5%14 / 21
ProgramBench: Can Language Models Rebuild Programs From Scratch?Coding
30.0%12 / 19
80.5% best5 / 19
APEX-AgentsAgentic
75.2% best1 / 13
37.8%6 / 13
SWE-MarathonCoding
43.8%
72.7%