Skip to main content
Breakdowns let you slice the same attributed activity across any dimension — person, tool, model, provider, or repository area — without losing the attribution thread that connects spend back to people and work. Every pivot stays anchored to real sessions and real changes, so the numbers you see are comparable across views.

​
By Person and Tool

The Person breakdown lists each engineer who has run at least one session in the selected period. For each person you see:
  • Sessions — total session count across all tools
  • Cost — total spend attributed to them
  • Cost per merged change — AI spend divided by the number of commits or PRs merged during the period
  • Entitlement usage — how their spend breaks down across seat subscriptions, metered API keys, and prepaid credits
The same engineer may appear under multiple tools if they use both Claude Code and Codex, for example. Each row in the tool sub-view is a separate harness, so you can see whether an individual’s spend is concentrated in one tool or spread across several.

​
By Model

The Model breakdown shows token spend and cost grouped by the model that handled each call. Models you might see include:
  • claude-sonnet-4.5
  • gpt-5-codex
  • opus-4.1
  • grok-4-fast
  • haiku-4
Calls to the same model made via different routes — for example, Claude Sonnet called directly through the Anthropic API versus routed through OpenRouter — are priced separately and labeled with their route. The model name alone does not determine the rate.
Use this view to understand whether your team is using the most cost-effective model for each task type, or whether expensive frontier models are being used in situations where a faster, cheaper model would produce equivalent results.

​
By Provider and Rail

The Provider breakdown groups spend by the organization that billed the tokens and by entitlement type: The rail column distinguishes between subscription seats (fixed cost per developer), metered keys (pay-per-token), and prepaid credits. Understanding rail distribution helps you decide whether to move more traffic onto seats or keep metered keys for overflow.

​
Change Context

From any row in any breakdown, you can navigate to the pull requests and commits associated with that person’s or model’s sessions. For each change Terma shows:
  • AI activity count — number of sessions that contributed to this change
  • Attribution score — confidence that the session output maps to the commit content
  • Review rounds — how many review cycles the PR went through before merge
  • CI retries — how many times CI had to re-run before the change landed
This context lets you connect spend to outcomes: a high-cost session that merged cleanly on the first review round looks very different from the same spend on a change that needed five review rounds and four CI re-runs.

​
Cross-Rail Findings

Terma surfaces findings when it detects spending patterns that suggest an inefficiency across your entitlement mix:
  • Rail imbalance — activity is heavily concentrated on one rail while another has available headroom (e.g. metered key spend is high while seat licenses sit underused)
  • Window throttles — engineers are hitting their seat rate limit by mid-day, forcing fallback to metered keys at higher per-token cost
  • Metered key handling work — tasks that required multiple long-context calls used a metered key when a seat subscription would have been cheaper for the volume
  • Personal keys on work repos — sessions attributed to engineers show API key spend that does not map to any team-provisioned entitlement, suggesting personal keys are in use for work tasks
Each finding links to the affected sessions and suggests a corrective action.

​
Cost per Merged Change

Cost per merged change is the headline efficiency metric in the Breakdowns view. It answers the question: for every piece of code that actually shipped, how much did AI assistance cost? You can compare this metric:
  • Across people — identify whether certain engineers get more value from the same AI spend
  • Across tools — compare whether Claude Code or Codex sessions result in a lower cost per merged change on your codebase
  • Over time — track whether the metric improves as your team builds better prompting habits and context documents

​
From the CLI

Use terma usage to pull the same data into your terminal. The --group flag controls the pivot dimension:
Group by model
Group by provider
Filter to one person
Output is a table by default. Add -o json or -o csv for machine-readable output suitable for further analysis or export to a spreadsheet.