§ / Benchmarks

Speed & cost, per recipe.

A modeled projection of each run's wall-time and marginal cost — container compute (instance vCPU/mem × active seconds) plus model inference for the model-calling runs. Numbers are generated from the same rate card that prices real executions; they are planning estimates, not a metered bill.

Modeled, not metered. These figures come from curated workload profiles priced through the rate card (re-check Cloudflare pricing before budgeting — see §06 Cost). The live, measured per-execution cost — real model tokens + modeled container compute — is on the Access-gated app dashboard (/analytics).

Recipe	Trigger	Instance	~Wall	Container	Model	~Cost / run
`pr-review`	event	standard-2	45s	$0.0002	$0.4200	$0.4202
`offload-test`	event	standard-2	8m	$0.0168	—	$0.0168
`matrix-fanout`	event	standard-2	8m	$0.0168	—	$0.0168
`vitest-shard`	event	standard-2	5m	$0.0105	—	$0.0105
`oxlint`	event	standard-2	25s	$0.0009	—	$0.0009
`cdp-acceptance`	event	standard-2	2m	$0.0042	—	$0.0042
`playwright-e2e`	event	standard-2	8m	$0.0168	—	$0.0168
`playwright-demo`	event	standard-2	2m	$0.0042	—	$0.0042
`product-demo`	cron	standard-3	3m	$0.0108	neurons*	$0.0108
`deploy-smoke`	event	standard-2	1m	$0.0021	—	$0.0021
`email-otp-login`	event	standard-2	1m30s	$0.0032	—	$0.0032
`refresh-fixtures`	cron	standard-2	2m	$0.0042	—	$0.0042
`ci-triage-pr`	event	standard-2	40s	$0.0002	$0.0585	$0.0587
`spec-drift-pr`	cron	standard-2	1m30s	$0.0010	$0.1800	$0.1810
`release-notes`	cron	standard-2	30s	$0.0002	$0.0540	$0.0542
`finops-audit`	cron	standard-2	30s	$0.0001	$0.0315	$0.0316

*neurons — the run's default backend is Workers AI catalog, billed as account-level Neurons with no per-call token usage, so no per-run USD is modeled. Switching pr-review and the other model-calling runs to the Anthropic or Bedrock backend bills per token at the model's rate (the Model column shows that metered figure where applicable). Hover a recipe name for its workload assumptions.