Benchmarks / Agent platform / Platform bench
Agent platform · Cost and start-upPlatform bench
What an agent costs at rest, what a sandbox that runs a shell costs to start cold, and what one call costs at the margin, each beside the figure its alternative publishes. Every script re-runs to produce its row.
Latest
History
3 results across 3 metrics, newest first. Superseded runs stay, so a revision's progress can be read.
| Measured | Subject | Metric | Value | Baseline | Result | Status |
|---|---|---|---|---|---|---|
| September 2026 | Hanzo | A dormant agent | 477 bytes | as its alternative publishes ~1 MB · 2,198× | Unscored | Frozen · Hanzo-measured |
| September 2026 | Hanzo | A sandbox that runs a shell, cold | 149.6 ms | as its alternative publishes <200 ms · ~1 s · 1–3 s · pooled 35.8 ms | Unscored | Frozen · Hanzo-measured |
| September 2026 | Hanzo | One call, marginal cost | $0.00000063 | as its alternative publishes $0.05 credit · 79,614× | Unscored | Frozen · Hanzo-measured |
Conditions
A dormant agent: 1M agents: 455 MB on disk vs ~977 GB · $0.01/mo vs $22/mo at list · 310,270 agents/s written · 0.031 ms resumeA sandbox that runs a shell, cold: A JavaScript context alone is 0.16 ms against a published 2.79 ms (17×) — a different primitive, quoted beside, not insteadOne call, marginal cost: ≈51 ms of compute per call · at a tenth of their price, 1B calls a month returns $3.7M after 25% to open source, three zones, a year of history and twice the infrastructure bill in peoplethe alternative figures are as their publishers state them, not re-run hereHardware
M-series laptop, September 2026
Source
hanzoai/benchmarks · fleet, sandbox, pricing
Reproduce
node fleet/fleet.mjs /tmp/fleet && node sandbox/sandbox.mjs && node pricing/pricing.mjs
More Agent platform benchmarks
Fleet residency · Live agent footprint · Sandbox cold start
Every result for Platform bench on the shelf · How these are measured