You're the person who has to decide how an AI agent reaches your company's systems. Your CRM holds the customers, a data warehouse holds the usage numbers, and an ITSM tool holds the support tickets. Someone on your team has asked you to pick an MCP gateway, and every pitch you've seen leads with demos. Nobody has shown you a price for getting the right answer.
That's the gap CData has just tried to fill. It published a benchmark of two MCP gateways serving the same business context to probabilistic agents. One is its own CData Connect AI Gateway. The other is an open-source MCP gateway it doesn't name. The study ran 810 runs by models from 12 developers, across three account-health tasks that touch a CRM, a data warehouse and an ITSM system.
The third task is the one you'll want to read twice. It was a write: add exactly the right 21 accounts to a review queue. CData's own words for the stakes are that "a wrong write lands in a live system."
The metric is the part you can steal. Cost per correct task is total spend divided by tasks done right, because "failed runs still cost money." Per-call pricing hides that. A cheap model that fails most of the time isn't cheap.
Here's what the numbers say. Through CData's gateway, 27 models reached correct answers with "about a tenth of the tool calls," at $0.26 per correct answer against $5.67 through the open-source gateway. The 21x headline rests on completion rates of 83% against 36%, and total spend of $89 against $828 over 405 tasks a side. The median run took 47 tool calls on the open-source gateway and 8 on CData's. Economy models completed 78% of tasks at $0.06 each on Connect AI and 17% at $1.48 on the other.
You should hold all of that at arm's length. These are a vendor's figures about its own product, measured against a competitor it chose not to name. Read them as a claim to test, not a verdict.
Two things survive that skepticism. First, CData published the benchmark harness, so you can rerun it with your own data. Few vendor benchmarks let you do that. Second, the mechanism makes sense. The open-source run lost because "the model did the work itself," paging and joining in its own reasoning. A gateway that runs the query costs fewer tokens than a model that rebuilds the join.
Now the turn. When you check the APIs.io catalog, CData's own record doesn't fully match the benchmark's third task. The provider page lists 14 API pages, led by the CData Connect AI MCP Server. The gateway's "one query across all three systems" is the CData Query API. Its system context comes from the CData Metadata API, and the live sources are defined in the CData Connection API.
The catalog scores CData at 83.8, in the exemplar band, with operational transparency at 92.1 and access clarity at 86.8. Contract governance, at 35.6, is its weakest facet. Its agent readiness is 67.3, which is high. Idempotency, delegated identity, error semantics and dynamic client registration are all lit, and that suits a gateway that executes queries for an agent.
Two dimensions are not lit: dry-run mode and reversibility documentation. Those are exactly what the write task needed. The agentic-access profile counts 76 operations, 43 of which act on data and 1 of which is flagged human-in-the-loop. CData defines a governed result as the agent writing "the right records and nothing it shouldn't have." Yet nothing in the published contract lets an agent preview a write or undo one.
So your takeaway is practical. Judge a gateway by cost per correct task, not cost per call, and count the failed runs. Rerun the published harness on your own systems before you believe the 21x. And when the task is a write, ask the vendor a question the benchmark didn't answer: can an agent preview what it is about to change, and can anyone undo it if it gets 21 accounts wrong?
Read the original: https://apis.io/2026/11/13/cdata-benchmarks-mcp-gateways-on-cost-per-correct-task/