Usage
Open Usage to inspect inference requests, input and output tokens, cache reads and writes, reasoning, and optional estimated cost. Tool steps and browser summaries do not add consumption of their own. Nothing is limited unless an administrator sets a usage limit.
Understand the numbers
Input/output totals, cache reads/writes and reasoning are shown only where the provider reports them. Reasoning and cache breakdowns are included in compatible totals; do not add them again. Unknown measurements stay unknown, and partial totals are labeled as known subtotals. Images and embeddings can have different or unavailable measurements.
Request details separate the configured model and provider, the returned model, and the provider that reported serving it. Fallback attempts appear individually. A provider-reported identity does not independently verify which physical model a gateway ran.
Cost estimates
Cost estimates help your organization monitor what its model usage costs. They are not an invoice or a charge from Decision Terrain, which licenses the platform separately. Estimates use customer rates when your administrators have saved them, and default public pricing otherwise. Customer rate estimates name their revision and effective date. Default public pricing can be stale and excludes negotiated terms, commitments and discounts. A bucket with no price leaves cost unavailable or incomplete. Estimates and provider-reported amounts are separate from invoices.
Usage limits
If an administrator has set a limit that an agent reaches, the agent stops starting new model requests and shows Usage limit reached with the time it resets. Requests already running finish, and messages and triggers sent meanwhile are kept and processed when the agent resumes. Ask an Owner, Admin or License Admin to raise or remove the limit if the work cannot wait. Most roles see only that status; limits are configured in License and Usage.
Get more from your usage
Start with representative work and a clear definition of success. Compare model choices using result quality, corrections, latency, request counts and known token consumption in Usage. Retries and fallback attempts count individually.
Reduce unnecessary repeated work, keep recurring schedules appropriate, and give agents concise, relevant context. Use request details to find expensive patterns. A stronger intelligence level can help with hard problems but does not guarantee a better or cheaper result, and it does not grant tools or bypass approvals.
When work is blocked
Check the timeline for approvals and missing credentials, agent active or paused state, current license and capability access, security or emergency restrictions, connected-provider readiness, and runtime limits. An ambiguous historical pause requires operator review.
See License and Usage for administrators.