Nightly read-only MCP audit: Haiku 5.5 vs Mistral Large 4 cost

For a nightly job that only reads (usage, balance, recent jobs), the model is the entire cost, and the Sume side is free to call. On the invented workload below, a 30-turn audit costs $0.02175 on Claude Haiku 5.5 and $0.27615 on Mistral Large 4, about 12.7 times more, or $0.65 against $8.28 over a 30-night month.

This works because OAuth with only mcp:read exposes read-only tools: the docs name jobs_list, assets_get, catalog_list and crawl_scrape as examples, and the tool inventory lists account_me, balance_get and usage_get among the read tools. Write and paid tools are hidden and return insufficient_scope if called. A runaway prompt therefore cannot spend generation credits, whatever model you pick.

Assumptions

The workload is made up for the arithmetic: 30 turns, 6,000 input tokens and 250 output tokens per turn, so 180,000 input and 7,500 output tokens per night. Prices are from the vendors' own pages, read 2026-10-09.

A read-only setup that matches the docs

Connect your client to the hosted endpoint, finish OAuth, and leave the Write toggle off on the consent page. Then check the session with mcp_health before you let the job loose.

• Call mcp_health and confirm authenticated.auth_source is mcp_oauth.
• Call tools_list; with Write off, only the read-only tools appear.
• Ask for balance_get, usage_get and jobs_list, then write the report. Stop there.

{
  "mcpServers": {
    "sume": { "url": "https://mcp.sume.com/mcp" }
  }
}

Limits of the comparison

Real audits grow their history each turn, so input tokens climb faster than the flat figure here; replace the assumptions with counts from your own logs. Mistral's page describes the preview API and open weights due by the end of October, and says nothing about MCP. Whether a Mistral client supports remote OAuth MCP is a client matter. Sume does not list either model as a selectable model in Agent Completions, which accepts only sume-agent.

Choosing between them

On these figures the audit is cheap either way: eight dollars a month is the high end. So the decision rests on other factors than price, such as which client your team already runs, whether that client supports remote MCP with OAuth, and how reliably the model follows a short report template. Anthropic's page also lists an adjustable effort setting for Haiku 5.5, which is a client-side way to trade cost against care on routine reads.

Because the job cannot spend generation credits, the worst case is a bad report, not a bad invoice. That is the argument for keeping unattended jobs on read-only OAuth and moving any paid step into a separate, supervised session with Write on and a max_spend_usd set.