SShogo Bench

DeepSeek V3.2

Hybrid-reasoning DeepSeek model with thinking and non-thinking modes, sparse attention, and tool-use

Shogo Agent Index

Supported#72 of 92
30.0range 22.1–37.8 · 27% of weight measured
Agentic coding
32
23% measured
Long-horizon autonomy
41
30% measured
Knowledge-work agents
24
32% measured
Computer use
18
estimated
Security & context learning
29
50% measured

Price

Input
$0.26/M
Output
$0.38/M
Cache read
$0.13/M
3:1 blend
$0.29/M
Agent-effective
$0.16/M

Cheapest host for agents: GMICloud at $0.060/M.

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

Agentic benchmarks

BenchmarkScoreNormalisedIndex weightHarness / source
SWE-bench Verifiedest.8%predicted 64
Terminal-Bench 2.039.6%8%Terminus 2
DeepSWEest.6%predicted 23
FrontierSWEest.5%predicted 4
CursorBenchest.4%predicted 26
GSO-Benchest.4%predicted 7
METR Time Horizonsest.10%predicted 63
Vending-Bench 2$1,0346%source
PostTrainBenchest.4%predicted 17
APEX-Agents21.3%8%source
GDPvalest.7%predicted 29
Remote Labor Indexest.5%predicted 1
DeepResearch Benchest.5%predicted 44
OSWorld 2.0est.6%predicted 2
OSWorldest.4%predicted 42
Cybenchest.5%predicted 44
CL-bench13.2%5%source

Other benchmarks

BenchmarkCategoryScoreSource
Aider PolyglotCoding74.2%source
ARC-AGIReasoning57.0%source
ARC-AGI-2Reasoning4.0%source
CritPtReasoning2.9%source
DTBenchKnowledge87.7%source
Fiction.LiveBenchLong context83.3%source
GPQA DiamondReasoning83.4%source
LMCAKnowledge29.1%source
OTIS Mock AIMEMath87.8%source
ProofBenchMath8.0%source
SciCodeCoding38.9%source
TheAgentCompanyKnowledge work42.9%source
WebDev ArenaCoding1367source

API hosts

Live prices and uptime for every host serving DeepSeek V3.2 via OpenRouter.

HostInputOutputCache readAgent $/MUptime 24hQuant
GMICloud$0.21$0.31$0.022$0.06099.7%fp8
Baidu$0.28$0.42$0.028$0.08099.8%fp8
AtlasCloud$0.26$0.38$0.13$0.1656.7%fp8
DeepInfra$0.26$0.38$0.13$0.1699.9%fp4
DigitalOcean$0.30$0.96$0.090$0.1698.6%—
SiliconFlow$0.26$0.80$0.14$0.19100.0%fp8