Once companies plug AI agents into their workflows, the bill is often the first thing that stops making sense: spending climbs, but nobody can say which agent, which model, or which request is driving it. On August 17, OpenRouter launched an Activity dashboard and a beta Analytics API aimed at exactly that problem.
The new tools break usage down by agent, model, and individual request. Managers can start from five top-line metrics — total spend, request count, token usage, cache hit rate, and blended cost per million tokens — and click through from there into the underlying logs. OpenRouter isn't promising to cut anyone's bill; what it's offering is a usage view that can be traced back to a point of accountability.
From a Total Down to a Single Request
OpenRouter's own documentation lists the dimensions usage can be sliced along: model, provider, API key, application, user, workspace, session, context length, and data region, among others. Charts can be aggregated by minute, hour, day, week, or month, or drilled all the way down to a single generation record.
Inside the logs, a single record can show the cost of inference, caching, web search, and file processing, alongside latency, throughput, time to first token, routing, and fallback information. For teams juggling multiple models and providers, these fields turn a vague "cost spike" into something you can actually investigate: was an agent's system prompt too long, did a cache prefix get broken, or did a routing switch push up latency?
Every company that spent the last two years deploying agents is now asking the same question: what are they costing us, and which ones are worth it?
That's how OpenRouter puts it. The dashboard exists to let that question land on a specific application and a specific request, rather than staying an abstraction.
Visibility Is Not the Same as Automatic Control
The product also includes a Guardrails view, which shows where prompt-injection and sensitive-data rules blocked, redacted, or flagged a request, and an Analytics API that lets teams pipe the same data into their own dashboards. What these tools solve is visibility — they don't set budgets or approval rules for a company.
That distinction matters. Once an agent is allowed to call tools, token cost is only one layer of the bill; the number of tool calls, the scope of retrieval, file processing, and the latency of external services all move the actual spend as well. Only after usage is sorted by agent, model, and request can a team meaningfully cap high-cost tasks or judge whether its caching and routing strategy is actually working.
OpenRouter's update moves AI usage analysis from a single lump-sum bill to a traceable stream of individual transactions. For engineering leads, the next step isn't admiring a tidy summary chart — it's finding which anomalous requests can be reliably reproduced, explained, and brought back within budget.
Sources: OpenRouter product announcement; verified via the Activity dashboard, Analytics API, and log fields; CocoLoop.