Latest release v2026.1023.0 is live (October 23, 2026).
See what's newAI spend should look like payroll, not like a surprise. Paperclip records the cost of every run against its task, warns you at a threshold, and stops new runs when a company, agent, or project reaches its cap. Enforcement runs on recorded cost, so runs already under way when the cap is crossed can take spend past it.
A budget in Paperclip is enforcement, not just a dashboard. At the warning threshold, 80% by default, Paperclip records a warning and the work continues. At 100% it opens a hard-stop incident, pauses the scope, cancels its queued and running runs, and refuses new ones. Tasks stay assigned and pick up again when the scope resumes. Spend is checked against recorded cost, which lands as each run finishes, so runs already under way can take the total past the cap. A budget set to warn only keeps spending.
Publish “Comparing agent runtimes” to the blog
Give a Researcher agent a $50 monthly budget. When its recorded spend this month reaches $40, Paperclip records a warning and the agent keeps working. When a finished run takes it to $50 or more, the hard stop opens: Paperclip cancels the agent’s queued and running runs and refuses new heartbeats, and its tasks stay assigned. If spend stood at $47 and the run that crossed the line cost $6, spend stands at $53 when the stop opens, plus whatever any other run in flight had used. A board user resumes the agent with Raise budget & resume and a new amount above current spend, or the budget resets on the 1st.
Each finished run records its provider, model, tokens, and cost against the run and its issue. The Costs page totals spend by agent and model, project, provider, and user for preset or custom date ranges, and an issue shows its own cost plus its sub-issues’. Subscription runs record $0; fees and provider invoices go in a separate Finance ledger.
Market brief: what changed since Monday
— Competitor X shipped usage-based pricing [source]
— Two enterprise threads mention agent budgets [source]
— Regulatory draft affects EU deployments [source]
3 claims verified · 1 refuted and omitted
Cost reporting helps you find expensive work. Langfuse records usage and cost for instrumented model calls, using supplied costs or model prices. Dashboards, alerts and its Metrics API help teams investigate spending and build controls. An alert still needs a person or another system to act on it.
Paperclip connects recorded spend to assigned work and the decision to start another run. With hard stops enabled, reaching a company, agent or project cap opens an incident and pauses that scope. It cancels queued and running runs and refuses new ones. Recorded cost can arrive after work has consumed it, so the run that crosses the cap and other work in flight can produce an overshoot. A board user can review the incident and raise the budget to resume; warn-only policies keep spending. See budget controls.
Request-path controls act at a different boundary. The LiteLLM proxy can check configured budgets for traffic routed through it. Those budget checks require a database; the docs warn that a database-free deployment does not enforce them. Traffic sent directly to a provider bypasses that proxy’s checks. Gateway budgets can have key or team scopes and reset windows; the gateway operator configures those limits and their recovery. Request-path enforcement also depends on the spend recorded when a request is checked; it does not make asynchronous accounting instantaneous.
Use reporting to understand spend, run budgets to decide whether agent work can continue, and a gateway when you need controls on routed model requests. The controls can work together. Check coverage and accounting delay before treating any limit as an exact invoice ceiling.
Each agent has its own adapter and model, set apart from its instructions, so you can staff mechanical work on inexpensive models and keep frontier models for judgment. The per-agent, per-model breakdown shows where the money goes; you decide where a cheaper model would do.
“AI spend should look like payroll: a number you set, a stop when it’s reached, and a receipt attached to every piece of work.”
Budgets, warnings, and receipts from the first agent you hire.