How to bring pricing, attribution, and enforcement to every AI agent in your enterprise
MuleSoft announces a host of new cost management capabilities for Omni Gateway, giving platform and FinOps teams the tools to see what AI actually costs and act on it: price every request against the rate you actually pay, cut the token waste agents never needed, cap budgets agents can’t overrun, and break spend down to the individual agent. Model Proxy is generally available now; further MCP Tool Optimizations, Cost Optimization for Agents & LLMs, Model Wallet, and the Unified Dashboard arrive later this quarter.
–
Let’s say you deployed a support agent that handled 40,000 conversations last month. What did it cost? The honest answer, for most enterprises, is a shrug and a promise to check a couple of dashboards.
Is it any wonder that 79% of large enterprises experienced AI cost overrun in the past year?
The business ships agents fast, each one reaching for the latest frontier model. Finance writes a number into the plan with no way to enforce it. Then the invoice lands and no one can trace which agent drove it.
The reason nearly four in five large enterprises overran their AI budget in the past year isn’t carelessness; it’s that spend often goes dark when agents start calling models directly. The priciest model answers by default even when a cheaper one would do, every team wires its own path, and a one-line question hits the same expensive endpoint as a deep reasoning task.
Omni Gateway, MuleSoft’s governance layer for both APIs & AI, already sits in the path of your Agent, MCP and LLM traffic. Implementing cost management tooling at the gateway means you can act on that traffic from a cost perspective, organized around actual usage, optimization strategies and visibility.
Which is why we’re introducing a wave of new AI cost management capabilities, built right into the gateway.
Start with a number you can trust
Model Proxy
When AI traffic runs through a proxy, everything needed for a real cost figure is captured in one place: the model that responded, the tokens it burned, and the rate your provider charges. Model Proxy sits in that path and prices every request against the rate you actually pay, so the figure on the dashboard is the figure on the invoice–no rate-card arithmetic, no estimated spend.
For agents running models outside the proxy, MuleSoft reports what it can verify: token volume and call counts.
From there, AI costs break down cleanly–spend can be rolled up from the single call to the agent to the business unit–so Support, Sales, and Marketing each get a chargeback figure that reconciles against the invoice.
Cut the wasted AI spend
Cost Optimization for MCP: definition tax and payload compression
MCP servers can burn hundreds of thousands of tokens per session, and most of it is definitional tax: full tool schemas shipped on every request when the agent touches one or two tools.
MuleSoft’s existing cost management dashboard already surfaces your heaviest MCP consumers, and two policies live at the gateway today cut the waste, one truncating bloated tool descriptions, the other filtering verbose tool output, without touching your servers or client code.
We’re introducing three new cost optimization policies for MCP:
- Progressive Disclosure stops shipping every schema upfront. The agent sees a lightweight list of tool names and pulls the heavy schema only for the tool it decides to use, so a two-tool task stops paying for a twenty-tool catalog.
- Smart Response Trimming reads the intent behind a call and cuts oversized tool responses down to what the agent actually asked for, so nothing downstream pays to carry fields it will never read.
- TOON compression rewrites unnecessarily weighty output; repeated JSON keys and tabular structures collapse into a compact format, applied only where it saves tokens and decoded back to JSON transparently. On list-heavy traffic that routinely takes off a third or more, invisibly to the calling agent.
Each of these ships as part of existing Cost Optimization for MCP capabilities, where MuleSoft turns MCP traffic from a token sink into a manageable expenditure.
Optimize all of your AI assets
Cost Optimization for Agents & Model Proxies
We’ve already outlined a host of available and upcoming Cost Optimization capabilities for MCP above, but MCP is just one part of a broader approach. Later this quarter we extend the same treatment to the two places the rest of your spend lives: agents and model proxies.
For Agents: MuleSoft reads your agentic workforce, surfaces the spend that shouldn’t be there, and delivers ranked recommendations that carry the saving, the confidence, and the fix.
For Model Proxies: the biggest lever is almost always right-sizing. A surprising share of premium-model traffic, classification, extraction, routine FAQ answers, would run on a cheaper tier and nobody would notice the difference. Our tools spot those workloads by reading task complexity and present the swap with the math done: projected monthly saving, expected quality impact, confidence behind both. Repeated prompts are another easy win, since a request that closely resembles one already answered never needs a model call: it gets served from the last answer, so you stop paying twice for the same question.
Across every one of these capabilities, the decision to implement stays with you, and FinOps gets a clean before-and-after record of every change.
Set a floor spend can’t cross
Model Wallet & Unified Dashboard
For everything the recommendations don’t catch, there’s a hard floor.
Model Wallet enforces budgets set per agent, per instance, or per business unit, tracked live as spend accrues. When a wallet runs down, enforcement kicks in before the overrun, not after the invoice explains it.
The Unified Dashboard expands our existing cost dashboard to include budget insights, so you have a single workspace where all of your AI cost management data resides. Actual spend, projected savings, and variance update continuously, and forecasting flags a projected overrun days ahead. Alert thresholds fire at the levels you choose, and the warnings reach you in app or via API while there’s still room to act.
The bottom line
Put these together and AI spend stops being a surprise. Understand agent costs, cut the waste before it compounds, right-size the spend that’s left, and set a floor nothing crosses, all from the gateway your traffic already runs through.
The enterprises that scale agents without flinching make cost a live signal, visible the moment it moves and manageable while there’s still something to do about it. Model Proxy is ready today. The rest lands over the coming weeks.
