# Hanzo — Benchmarks > What the platform costs and what it scores: 477 bytes to keep an agent and 125 ns to start one, the engine against llama.cpp by prompt length, the router on GPQA-Diamond against the field, and Hanzo Context on LoCoMo, MemoryAgentBench, LongMemEval and RepoBench-R — with the losses and the gaps. SYSTEMS BENCHMARKS # 98.0% on GPQA-Diamond, and What It Costs Tiered dispatch reaches frontier accuracy at a fraction of frontier price. Below it: what an agent costs at rest and in flight, how fast a sandbox starts, and what retrieval recovers — with the losses and the gaps. [Explore benchmarks ↓](#sec-routing)[Reproduction protocol →](#sec-proof) [477 BDormant state residency →](https://hanzo.ai/benchmarks/fleet)[125 nsGoroutine process spawn →](https://hanzo.ai/benchmarks/goroutine)[0.13 msV8 context allocation →](https://hanzo.ai/benchmarks/sandbox)[98.0%GPQA-Diamond accuracy →](https://hanzo.ai/benchmarks/gpqa)[37.5LoCoMo multi-hop ALL@20 →](https://hanzo.ai/benchmarks/locomo)[67 / 100MemoryAgentBench 262K →](https://hanzo.ai/benchmarks/memoryagentbench)[33.2RepoBench-R code R@1 →](https://hanzo.ai/benchmarks/repobench-r)[1.088×CUDA throughput vs llama.cpp →](https://hanzo.ai/benchmarks/inference) 01   MODEL ROUTING ## Pareto Frontier on GPQA-Diamond Reasoning accuracy versus token pricing across tiered model dispatch. [Detailed GPQA-Diamond evaluation table →](https://hanzo.ai/benchmarks/gpqa) Model Accuracy (%) Correct / Total Cost (per 1M) Enso Ultra 98.0 194 / 198 $25 Enso 96.0 190 / 198 $20 Enso Flash 92.9 184 / 198 $4 GPT-5.6 Sol 92.9 — $25 Dynamic Cascade Dispatch Speculative verification on high-difficulty reasoning benchmarks. The Flash routing tier achieves 92.9% accuracy at 84% lower token expenditure ($4/M vs $25/M) through selective fallback. ROUTING TIERS: ULTRA · STANDARD · FLASH [Detailed GPQA routing evaluation →](https://hanzo.ai/benchmarks/gpqa)Related:[GPQA-Diamond](https://hanzo.ai/benchmarks/gpqa)[Enso Router](https://hanzo.ai/enso)[Inference throughput](https://hanzo.ai/benchmarks/inference) [← 03 Cold Start & Isolation](#sec-start)|[04 Memory & Context →](#sec-memory) 02   RUNTIME AND CONCURRENCY ## Process Concurrency & Memory Residency Heap residency, serialization throughput, and resumption latency under concurrent agent execution. [477 bytesDormant process stateSerialized agent snapshot on disk; sub-millisecond resumption on wake.](https://hanzo.ai/benchmarks/fleet)[601 bytesHeap per live agentActive goroutine memory footprint per concurrent agent process.](https://hanzo.ai/benchmarks/goroutine)[309,789/secState write throughputSustained disk write throughput under batch state serialization.](https://hanzo.ai/benchmarks/fleet)[0.034 msIn-process resumptionMemory unmarshal latency from disk snapshot to running goroutine.](https://hanzo.ai/benchmarks/goroutine) [Detailed fleet residency benchmark →](https://hanzo.ai/benchmarks/fleet)Related:[Goroutines](https://hanzo.ai/benchmarks/goroutine)[Fleet residency](https://hanzo.ai/benchmarks/fleet)[Sandbox isolation](https://hanzo.ai/benchmarks/sandbox) [03 Cold Start & Isolation →](#sec-start) 03   STARTUP AND ISOLATION ## Execution Sandbox Cold-Start Latency Initialization latency across V8 contexts, V8 isolates, and microVM instances without prewarmed standby pools. [Detailed sandbox isolation benchmark →](https://hanzo.ai/benchmarks/sandbox) Process and Sandbox Spawn Latency Live agent (goroutine) 125 ns V8 context 0.13 ms V8 isolate 0.35 ms MicroVM boot 309 ms [125 nsGoroutine spawn →](https://hanzo.ai/benchmarks/goroutine)[0.13 msV8 context allocation →](https://hanzo.ai/benchmarks/sandbox)[0.35 msV8 isolate instantiation →](https://hanzo.ai/benchmarks/sandbox)[309 msMicroVM kernel boot →](https://hanzo.ai/benchmarks/sandbox) [Detailed sandbox isolation benchmark →](https://hanzo.ai/benchmarks/sandbox)Related:[Sandbox isolation](https://hanzo.ai/benchmarks/sandbox)[Goroutines](https://hanzo.ai/benchmarks/goroutine)[CUDA throughput](https://hanzo.ai/benchmarks/inference) [← 02 Runtime & Concurrency](#sec-runtime)|[01 Model Routing →](#sec-routing) 04   MEMORY AND RETRIEVAL ## Multi-Hop Retrieval & Fact Reconciliation Recall accuracy and temporal conflict resolution over extended context windows (up to 262K tokens). [Detailed LoCoMo multi-hop benchmark →](https://hanzo.ai/benchmarks/locomo)[Detailed MemoryAgentBench 262K table →](https://hanzo.ai/benchmarks/memoryagentbench) [LoCoMo multi-hop retrieval → Baseline Hanzo18.8 → 37.5ALL@2018.837.570.9 → 80.7R@2070.980.7](https://hanzo.ai/benchmarks/locomo)[MemoryAgentBench →Temporal conflict resolution at 262K context33 → 673367](https://hanzo.ai/benchmarks/memoryagentbench) [Detailed LoCoMo memory benchmark →](https://hanzo.ai/benchmarks/locomo)Related:[LoCoMo](https://hanzo.ai/benchmarks/locomo)[MemoryAgentBench](https://hanzo.ai/benchmarks/memoryagentbench)[LongMemEval](https://hanzo.ai/benchmarks/longmemeval)[Subject Scope](https://hanzo.ai/benchmarks/locomo-subject-scope) [← 01 Model Routing](#sec-routing)|[05 Code Retrieval →](#sec-code) 05   CODE RETRIEVAL ## AST-Aware Cross-File Dependency Retrieval Syntactic and dependency graph traversal for code completion on RepoBench-R. [Detailed RepoBench-R code retrieval table →](https://hanzo.ai/benchmarks/repobench-r) [RepoBench-R →Cross-file context retrieval for code completion.18.8 → 33.2R@1 · +14.4 pts18.833.2](https://hanzo.ai/benchmarks/repobench-r)Dependency Graph Traversal vs Lexical Baseline Lexical BM25 retrieval degrades across multi-file repositories due to symbol shadowing. Semantic graph indexing walks import hierarchies and call trees, boosting R@1 from 18.8 to 33.2 (+14.4 points over baseline). HARNESS: REPOBENCH-R (PYTHON) · R@1 METRIC [Detailed RepoBench-R code retrieval benchmark →](https://hanzo.ai/benchmarks/repobench-r)Related:[RepoBench-R](https://hanzo.ai/benchmarks/repobench-r)[Hanzo Dev](https://hanzo.ai/dev)[LoCoMo](https://hanzo.ai/benchmarks/locomo) [← 04 Memory & Context](#sec-memory)|[06 Inference Speedup →](#sec-inference) 06   INFERENCE THROUGHPUT ## Prefill & Decode Throughput vs llama.cpp Throughput evaluation of Hanzo Engine against llama.cpp on identical hardware, weights, and quantization levels. [Detailed cross-backend inference matrix →](https://hanzo.ai/benchmarks/inference) Backend Model Context Speedup CUDA Zen-1.7B pp4096 1.088× CUDA Zen-4B pp2048 1.051× Measured under identical FP16/Q4_K quantization on NVIDIA Hopper and Ada architectures. Fused Attention & Paged KV-Cache FlashAttention-3 integration and custom fused GEMM kernels prevent memory bandwidth saturation during prefill. Paged KV-cache allocation eliminates memory fragmentation under concurrent batch requests. BACKENDS: CUDA (SM_90/SM_89) · METAL (APPLE SILICON) · AVX-512 [Detailed cross-backend inference matrix →](https://hanzo.ai/benchmarks/inference)Related:[Inference matrix](https://hanzo.ai/benchmarks/inference)[Startup sandbox](https://hanzo.ai/benchmarks/sandbox)[Zen models](https://hanzo.ai/zen) [← 05 Code Retrieval](#sec-code)|[07 Evaluation Protocol →](#sec-proof) 07   REPRODUCIBILITY & PROTOCOL ## Evaluation Protocol & Reproduction All benchmark harnesses, raw run telemetry, and analysis scripts are published in the open repository. Every published figure is directly regenerable from committed test fixtures. Command Line Reproduction Clone the evaluation harness and execute the test runner to regenerate raw telemetry and comparison tables from scratch: Terminal:`git clone https://github.com/hanzoai/cloud && cd cloud/bench/brain && make reproduce` Deterministic Execution Invariant greedy decoding (temperature 0.0), fixed pseudo-random seeds, and uniform prompt templates across all evaluation arms. Bootstrap Confidence Intervals 95% bootstrap confidence intervals computed over questions and test splits to distinguish empirical signal from stochastic variance. Hardware Isolation Bare-metal benchmarking with pinned CPU affinity, isolated memory regions, and dedicated GPU/accelerator contexts. Related:[Harness repository](https://github.com/hanzoai/cloud/tree/main/bench/brain)[Research papers](https://papers.hanzo.ai) [← 06 Inference Throughput](#sec-inference)