AI Latency Benchmarks

Round-trip latencies from AWS Lambda to sandbox providers and LLM inference APIs.

Results

Methodology

• Lambda: Node.js 20.x, 3072 MB, us-east-1 (Ashburn, VA)
• Samples: 100 per test (10 fresh sandboxes x 10 samples each)
• Sandbox tests: Coldstart = create() + delete(). Tool execution = echo ok via provider SDK. Network round trip = Sandbox Agent GET /v1/health.
• LLM tests: Minimal prompt ("Hi", max_tokens=1) measuring network + inference overhead, not generation speed.
• Sandbox provisioning and agent installation are excluded from exec/health measurements. Sandboxes are reused across samples within each iteration.

Network topology

	ASN	Location
Lambda	AS14618 Amazon.com	Ashburn, VA
E2B	AS396982 Google LLC	The Dalles, OR
Daytona	AS30633 Leaseweb USA	Centreville, VA

Usage

# Run all tests
./invoke.sh /run "tests=*&samples=5"

# CSV output
./invoke.sh /run "tests=llm:*&samples=10&format=csv"

# Full benchmark (10 iterations x 10 samples, outputs sandbox-raw.csv + provider-raw.csv)
./bench.sh

# Network info from Lambda
./invoke.sh /netinfo

# Deploy
./deploy.sh