OpenAI charges no per-call MCP fee, but Sume paid tools bill

Using a remote MCP tool in the Responses API adds no per-call fee from OpenAI. Its guide says you pay for the tokens used when importing tool definitions or making tool calls, and no additional fees apply per tool call. A Sume paid tool is a separate charge: the generation spends credits from your Sume balance.

So a single agent turn has two meters. Tokens go on the OpenAI bill. Image, video, music and speech jobs go against the Sume wallet. Treat the second one as the risk, because a model can request many jobs quickly.

Two meters, one agent turn

Which Sume fields cap a paid call?

Paid and write calls need an idempotency_key. dry_run=true returns an admission and cost preview and does not submit the job. max_spend_usd is optional, and Sume enforces it only when you send it (tools and gates). Admission can also refuse with 429 queue_full or rate_limited; honor retry-after (generation admission).

Do not read the OpenAI sentence as 'MCP calls are free'. It describes OpenAI's fee, not the server's.

A safe default

• Always run dry_run first and show the estimate to the user.
• Send max_spend_usd on every paid call.
• Reuse one idempotency_key per intended job, so a retry cannot buy twice.
• Keep paid tools behind require_approval.