Say you run the platform team that sits between your company's agents and the MCP servers they call. Last month someone asked you to bring token costs down. You've read the vendor claims, you've seen big percentages, and now you have to decide which gateway policies to switch on. A number on a slide won't help you with that. You need to know what was measured, and how far it can be trusted.
MuleSoft has published the methodology behind its MCP cost claims, which is rarer than the claims themselves. In a post titled MCP Cost Management: Inside Our Benchmarks, its team explains how it measured five gateway policies that shrink what MCP servers send to agents: progressive disclosure, smart response trimming, tool mapping, compressing repeated structures and cleaning payloads. It ran them against 18 public MCP servers, from Notion and Redis to Jira, GitHub, Stripe and arXiv, through its Omni Gateway. Each policy was compared against a no-policy baseline. These are MuleSoft's own figures.
Here is the first thing you'd notice: the savings vary enormously by server. Progressive disclosure hides the tools a task doesn't need. It cut 91.8 percent of MCP tokens on Notion and 87.6 percent on Redis. On arXiv it cut nothing. Smart response trimming went the other way, cutting 93.9 percent on arXiv and 60.4 percent on Stripe. Apify, whose own measurements this series covered two weeks ago, saw 56.7 percent from progressive disclosure. So whichever policy you turn on, the result depends on which servers your agents actually use.
The second thing is what the post says the numbers don't prove. The percentages "cannot be directly added together," so you can't stack five policies and total up the savings. They measure only MCP tokens at the gateway, not the agent's own prompts and reasoning. And they "do not establish equivalent reductions in a complete agent workflow or prove unchanged answer quality." If you were about to forecast a budget from the headline figures, that last caveat is the one to keep in front of you.
The method story may be the most useful part. MuleSoft's first benchmark ran against live servers, and it failed because agents took different numbers of turns. In the post's words, "a run with one extra turn can consume substantially more tokens even when the policy reduces the individual payloads." So the team froze responses into a golden set and judged answer quality separately. If you ever test a policy on your own estate, that's worth copying: pin the inputs first, or turn count will swamp what you're trying to measure.
Now the turn about the company's own record. The catalog holds the platform this runs on. MuleSoft's provider page lists API pages including the MuleSoft Anypoint Platform and the MuleSoft Anypoint Exchange API, where the MCP servers a gateway governs are catalogued. On the Agent Readiness score, a well-known catalog, an agent card, delegated identity and protected resource metadata are lit, a rarer set than most of the catalog has. The Kin Score is 48.3, in the developing band, with contract quality at 65.2 and developer ergonomics at 65.5, against access clarity at 25.0. The Agent Readiness score is 43.4. And the MCP server dimension is still unlit on the record of a company publishing benchmarks of other people's MCP servers, as it was when this series covered MuleSoft's open policies last week. That doesn't undo the benchmarks. It does mean the company measuring everyone else's servers hasn't lit that dimension for itself.
What you take away is fairly simple. Treat any single-policy percentage as a result for one server, not a promise for yours. Don't add the figures together. Don't read gateway token savings as savings for the whole agent workflow. And the rule MuleSoft arrives at, that turn count dominates per-call savings, is the same one TestMu AI found from Slack's data this week. Two vendors, two methods, one finding: an agent's cost is the task, not the call. So when you plan your cost work, measure the task end to end, and hold answer quality fixed while you do.
Read the original: https://apis.io/2026/10/23/mulesoft-benchmarks-mcp-cost-policies-and-says-what-the-numbers-do-not-prove/