Costs

The Costs surface (dashboard route /dashboard/costs, under the Observe chapter) is your cost command center. It rolls every span's USD cost into headline spend, surfaces heuristic recommendations for where money is leaking, and lets you slice spend across eight breakdowns. When you need to know what you spent, who spent it, and what to cut, this is where you start.

Top controls

Two controls sit above everything:

• Range selector. Choose 24h, 7d, 30d, or This month. Every card and tab below recomputes for the selected window.
• Saved Views. Switch between prebuilt views and your own. Views persist through the API, so a saved slice is there the next time you open the surface.

Headline cards

At the top you get an estimated blocked spend card and a budget progress card, followed by four stat tiles:

• Total Spend for the range.
• Input Tokens.
• Output Tokens.
• Savings landed, which highlights green when it is above zero.

Below the tiles, a heuristic Recommendations list flags patterns worth investigating, such as reflection loops, oversized models, and cost outliers.

The breakdown tabs

The lower half of the surface is an eight-tab breakdown. The active tab is selected by the ?breakdown= URL param. Overview is the default and carries no param.

Overview

A composite view that stitches the other tabs together: a spend-over-time trend, a model x team cost matrix heatmap, an Outcomes summary, top agents by cost, and a "Savings opportunities" list of open cost findings. Each panel deep-links to its dedicated tab or surface, so Overview is both a dashboard and a launchpad.

Outcomes

Cost framed against results: cost per successful run, success rate, and the wasted spend on failed traces. A provenance badge shows how many outcomes are explicit versus proxy-derived, so you know how much of the picture is measured versus inferred.

By Dimension

Cost grouped by a custom dimension key that you attach to events. Use it to roll spend up by a label that matters to your business (a tenant, a feature flag, a customer tier) rather than the built-in dimensions.

Trend

Spend over time with selectable granularity (hour, day, week, or month) and an optional group-by series, so you can watch a single line or split it into bands.

Per-run Cost

The per-run cost distribution plus a "cost composition by event kind" table that splits each run into LLM calls, tool calls, and governance decisions. Use it to find the runs that cost the most and what they spent it on.

Explorer

Ad-hoc slicing. Pick a group-by dimension (Model, Agent, Session, Environment, or any custom key), apply optional preset chips (Model, Agent, Session), and toggle the sort between By Cost and Biggest Mover. Adding a second "vs" dimension renders a cross-dimension cost matrix heatmap. A ?preset= URL param deep-links a starting dimension.

Delegation

Spend grouped by the chain of agents that called each other (the delegation path). This is how you see which agent collaborations cost the most, rather than which single agent does.

Reports

Scheduled cost reports. Each report carries a name, a kind, a cadence, a delivery target, and an enable toggle.

Getting here from the home dashboard

The home dashboard shows a SpendGlance summary (total spend plus a sparkline) that links straight to this surface once a project has traces. If you have not sent any traces yet, follow the Quickstart first.

Related surface: Threads

When you want spend grouped by conversation rather than by dimension, open Threads (/dashboard/threads). It groups every trace that shares a session_id into a multi-turn conversation and shows the cumulative cost across the whole exchange.

Next steps