SShogo Bench

Qwen3.5 397B-A17B

Large open Qwen multimodal MoE for visual agents and long technical tasks

Shogo Agent Index

Unranked

Not ranked yet: 1 agentic benchmark measured, and we need at least 2 plus an overall capability score. It will be ranked as results are published.

Agentic coding
37
estimated
Long-horizon autonomy
44
estimated
Knowledge-work agents
27
32% measured
Computer use
24
estimated
Security & context learning
37
estimated

Price

Input
$0.60/M
Output
$3.60/M
Cache read
—/M
3:1 blend
$1.35/M
Agent-effective
$0.74/M

Cheapest host for agents: DeepInfra at $0.31/M.

Agent-effective = 20 input tokens per output token, 85% cache hits. Why.

Agentic benchmarks

BenchmarkScoreNormalisedIndex weightHarness / source
SWE-bench Verifiedest.8%predicted 69
Terminal-Bench 2.0est.8%predicted 45
DeepSWEest.6%predicted 30
FrontierSWEest.5%predicted 6
CursorBenchest.4%predicted 29
GSO-Benchest.4%predicted 12
METR Time Horizonsest.10%predicted 67
Vending-Bench 2est.6%predicted 20
PostTrainBenchest.4%predicted 19
APEX-Agents24.9%8%source
GDPvalest.7%predicted 35
Remote Labor Indexest.5%predicted 2
DeepResearch Benchest.5%predicted 46
OSWorld 2.0est.6%predicted 4
OSWorldest.4%predicted 54
Cybenchest.5%predicted 55
CL-benchest.5%predicted 18

Other benchmarks

BenchmarkCategoryScoreSource
CritPtReasoning0.9%source
DTBenchKnowledge87.5%source
FrontierMath Tiers 1-3Math31.2%source
GPQA DiamondReasoning86.4%source
LMCAKnowledge37.9%source
OTIS Mock AIMEMath88.9%source

API hosts

Live prices and uptime for every host serving Qwen3.5 397B-A17B via OpenRouter.

HostInputOutputCache readAgent $/MUptime 24hQuant
DeepInfra$0.36$2.40$0.18$0.31100.0%fp8
StreamLake$0.60$3.60$0.12$0.3597.5%—
DeepInfra$0.45$3.00$0.22$0.3996.8%fp8
Alibaba$0.39$2.34—$0.48100.0%—
Parasail$0.50$3.60$0.30$0.4998.9%fp8
DeepInfra$0.68$4.50$0.33$0.5877.7%fp8
AtlasCloud$0.55$3.50$0.55$0.69100.0%fp8
GMICloud$0.60$3.60—$0.74100.0%fp8
Novita$0.60$3.60—$0.74100.0%—
Venice$0.75$4.50—$0.9396.9%—