Clef costs $0.24 per million input tokens: what gating a batch costs

Gating with Clef costs $0.24 for every million input tokens, and Cloudflare's model page says output tokens are not charged. So the gate's price is your total input tokens times 0.24 divided by one million; there is no second term. The cost to Sume is separate and is shown by a dry run.

The arithmetic

This table uses only the vendor rate and the vendor's per-image cap. Your own brief lengths are the unknown, so measure them with the usage field that the response carries.

Compare it with the thing you are guarding

A Sume paid call tells you its cost before you commit. On hosted MCP, dry_run=true returns the estimate as usage.credits and does not submit; the tools and gates page documents it. If a dry run is free and exact, a Clef gate earns its place only where it spares a person's time or screens things a dry run cannot see, such as whether the brief makes sense.

Record usage for each call for a week, then multiply. Do not assume a brief's token count.

Where the gate pays and where it does not

A gate has three costs: the vendor charge, the added latency and the chance it blocks something good. Cloudflare's blog reports a median latency of 38.8 ms for Clef-flash and 209.3 ms for Clef, so for a batch job the time is small next to a render.

The saving comes from what you do not submit. A brief that fails the gate never reaches Sume, and a brief that passes still gets a dry_run. Count blocked briefs for a week. If the gate blocks nothing, drop it; if it blocks good briefs, loosen the threshold.

• Log every block with the state you sent.
• Review a sample of blocked briefs.
• Treat the probability as advice, not a cap.