Back to model catalog

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.

Hy4 flashHy4Hy3Chat with these 3
AttributeHy4 flashtencent/hy4-flashHy4tencent/hy4Hy3tencent/hy3
Specs
Context window1M1M262K
Max output66K66K131K
Input modalitiesTextImageTextImageText
Output modalitiesTextTextText
Capabilities
StreamingSupportedSupportedSupported
Tool callingSupportedSupportedSupported
Structured output (JSON)SupportedSupportedSupported
Image inputSupportedSupportedNot supported
Thinking modeSupportedSupportedSupported
Price / 1M tokens
Input$ 0.080$ 0.96$ 0.16
Output$ 0.32$ 2.88$ 0.64
Cache read$ 0.0032$ 0.048$ 0.040
Cache write$ 0.080$ 0.96$ 0.16
What it is good at
Best for
  • Documents and slides
  • High-volume everyday tasks
  • Images and computer use
  • Long-context work
  • Agentic coding
  • Images and computer use
  • Finance and office work
  • Long-context work
  • Agentic coding
  • Productivity agents
Published benchmarks
Terminal-Bench 2.1pass_rate——
  • 71.7%effort high · Terminus-2(setup not fully disclosed)
Terminal-Bench 4.0pass_rate
  • 19.2%effort high · Claude Code(setup not fully disclosed)
  • 44.4%effort high · Claude Code(setup not fully disclosed)
—
BrowseComp (with tools)accuracy——
  • 84.2%effort high(setup not fully disclosed)
Claw-Eval Generalpass3_rate——
  • 68.5%effort high(setup not fully disclosed)
AutomationBench v1.0.6score
  • 52.2%effort high(setup not fully disclosed)
  • 56.2%effort high(setup not fully disclosed)
—
DeepSWE 1.1resolved_rate
  • 64.6%effort high · mini-swe-agent(setup not fully disclosed)
  • 74.9%effort high · mini-swe-agent(setup not fully disclosed)
—
ProgramBenchscoresetups differ by: harness
  • 16%effort high · mini-swe-agent(setup not fully disclosed)
  • 30%effort high · Claude Code(setup not fully disclosed)
—
SWE-Bench Multilingualresolved_rate——
  • 75.8%effort high · SWE-agent(setup not fully disclosed)
SWE-Bench Proresolved_rate——
  • 57.9%effort high · SWE-agent(setup not fully disclosed)
SWE-Bench Verifiedresolved_rate——
  • 78%effort high(setup not fully disclosed)
OSWorld 2.1 (partial credit)score
  • 68.1%effort high(setup not fully disclosed)
  • 78.9%effort high(setup not fully disclosed)
—
AA-LCRscore——
  • 73.4%effort high(setup not fully disclosed)
MMMU Proaccuracy
  • 75.8%effort high(setup not fully disclosed)
  • 78.6%effort high(setup not fully disclosed)
—
ZeroBench main (with tools)pass_at_5
  • 45.8%effort high · 10 runs(setup not fully disclosed)
  • 55.4%effort high · 10 runs(setup not fully disclosed)
—
Humanity's Last Exam (with tools)accuracy——
  • 53.2%effort high(setup not fully disclosed)
Humanity's Last Exam Full (with tools)accuracy
  • 55.2%effort high(setup not fully disclosed)
  • 59.1%effort high(setup not fully disclosed)
—
GPQA Diamondaccuracy——
  • 90.4%effort high(setup not fully disclosed)
Terminal-Bench-Science 0.1accuracy—
  • 19.8%effort high · Claude Code(setup not fully disclosed)
—
MCP Atlas Publicscore——
  • 79.1%effort high(setup not fully disclosed)