Spend tells you what every agent really costs.

AI cost attribution,
before the bill lands.

Provider dashboards show tokens by API key. Agent Ping shows cost by agent, by customer, by feature. Priced on our servers from a maintained price list, and waiting in your dashboard long before the invoice lands.

Start free All features
spend · this month↑ on budget
$5,382
spent
$0.094
cost / successful run
content-writer$2,189
research-agent$1,474
support-triage$685
email-classifier$262

attribution

Cost, broken down by who actually ran it.

Tag a run with a customer and a feature once. From then on, every token is attributed, and the rollup is waiting in the dashboard whenever finance asks who burned what. If the term is new, what is AI cost attribution? covers it without the product.

  • Per-agent, per-customer, per-feature cost rollups
  • Tag customer_id and feature at run start; rollups are retroactive once tagged
  • Unattributed runs flagged, so the misconfigured cron stops hiding
spend · top customers by agentcost / run
AgentCustomerCost/ run
support-triageacme-corp$1,460$0.118
support-triagezenith-labs$786$0.096
support-triage(unattributed)$358$0.585
content-writeracme-corp$967$0.310

the metric finance trusts

Budget on cost per successful run.

The metric finance can actually budget on: your total spend over the runs that finished cleanly. Runs that failed halfway burned money too, and this is the number that counts it. Priced on our servers, so it lines up with the invoice.

  • Server-side pricing per provider and model, override any row per team
  • Cache-aware: prompt-cache reads priced apart from fresh input
  • Priced server-side, so your totals line up with the provider bill
this month↑ on budget
$5,382
total spent
$0.094
cost / successful run
content-writer$2,189
research-agent$1,474
support-triage$685

anomaly alerts

Catch the runaway before the invoice.

Every agent learns its own baseline. The day one breaks it, the alert reaches you, not your accountant three weeks later.

  • Anomaly detection on a 14-day per-agent baseline
  • Today more than 2.5σ above the trailing mean fires the agent's route
  • Alerts to Slack, email, or any webhook you already watch
spend alert↑ 3.1σ
research-agent spend up 280% today $409 vs $108 baseline · (unattributed) runs · sent to #ai-ops
content-writer nearing monthly budget 82% of $2,500 cap · 6 days left in cycle

spend guard

Alerts report the runaway. Guard refuses the next one.

Every observability tool can tell you a loop burned the budget overnight. Guard is the opt-in brake: one call at the top of a scheduled script asks "am I clear to run?" before the expensive work begins. Trip a spend rule, or hit pause in the dashboard, and the next run simply does not start.

  • Rules on total spend, spend in a window, or run count, keyed by customer, agent, and function
  • Checked against your real spend on our side, no counters to keep in your own code
  • Remote pause switch: stop an agent from the dashboard, no deploy
  • Fail closed by default, fail open per agent if you prefer (Team and Business)

Guard docs →

guard · before the run starts1 blocked
  • acme-corp · daily cap $200
  • nightly-summariser · runs / day 40
  • research-agent paused
nightly-summariser refused to start acme-corp at $214 of $200 daily cap · runs again when you clear it

No more month-end surprises.

The details that turn a token bill from a monthly surprise into a line item you can forecast.

pricing

Rate card per model

Defaults ship for every Anthropic and OpenAI model; override any row for your negotiated rates.

accuracy

Cache-aware accounting

Prompt-cache reads are priced apart from fresh input, so a heavily-cached agent does not look far more expensive than it really is.

hygiene

Unattributed detection

Runs you never tagged with a customer have nowhere to put their cost. We surface them, so the forgotten cron racking up a bill cannot hide.

guard

A brake, not just a gauge

Opt in per script and a tripped spend rule refuses to start the next run. Pause any agent from the dashboard, no deploy. Team and Business.

breakdown

Where the money goes

Cost by provider and model across Anthropic, OpenAI, Bedrock, and Gemini in one view.

control

Budget alerts

Set a monthly cap per agent and hear about it at 80%, not when the card declines.

Cost, end to end.

Attribute every token

Tag once and every run rolls up by agent, customer, and feature, retroactively.

  • acme-corp $2,427
  • zenith-labs $1,332
  • (unattributed) $767

The number the invoice will show

We apply the prices on our side from a maintained rate card, so a stale SDK can never report the wrong cost.

$0.094
cost / run
57,359
runs

Catch the spike early

Each agent's own baseline means a quiet runaway gets loud before the invoice does.

research-agent +280%3.1σ over baseline

See which customer is eating the margin.

Spend broken down by the dimensions your business runs on, not by API key. One customer is usually well ahead of the others, and on flat-rate pricing that is the one being served at a loss.

spend · by customer · last 30 days $5,382 total
CustomerCostRunsCost / runChange
acme-corp$2,42718,204$0.133↑ 214%
zenith-labs$1,33226,880$0.050± 3%
globex$86117,410$0.049± 1%
northwind$3557,120$0.050↓ 6%
unattributed$767———

acme-corp costs 2.7× per run what everyone else does. Open any run and the reason is on it.

run · nightly-report-agent success
Workflowweekly-digest
Customeracme-corp
Featureexec-summary
Provideranthropic
Modelclaude-opus-4-8
Tokens21,904
Cost$0.21
Latency38.6s
Retries0
Alert statenone
Quality signal4.6 / 5

Questions, answered.

How does Agent Ping know what each agent costs?
Every LLM call your agents make is priced server-side from a rate card per provider and model, so the number never depends on the SDK guessing. Tag a run with customer_id and feature and the cost rolls up by agent, customer, and feature automatically.
What is AI cost attribution?
AI cost attribution is answering who spent the money: which agent, which customer, which feature, which model. A provider dashboard shows tokens by API key; Agent Ping prices every LLM call on our servers and ties it to the run that made it, so the bill splits along the lines your business actually uses.
Can I see LLM cost broken down by model?
Yes. Every llm_call event carries the exact provider and model string, and the Spend page groups cost by agent, model, customer or feature. A mid-month swap to a pricier model shows up as its own line, and any model missing from your rate card is listed as unpriced rather than silently counted as free.
How do I attribute token cost per user or per customer?
Pass customer_id when a run starts. It takes any string, so it can be an account, a workspace or an individual end user, whichever level you bill or budget at. Cost then rolls up per customer in the dashboard, and you can see which accounts cost more to serve than they pay.
How do I split AI spend between internal teams?
Two ways. Tag runs with a feature that maps to the owning team, or give each team its own Agent Ping team with its own API key, budgets and alert routes. Teams are free on every plan and seats are unlimited, so separating spend by department costs nothing extra.
Do I have to change my code to get cost attribution?
Two lines to initialise the SDK, then it auto-captures tokens on every Anthropic, OpenAI, Bedrock, and Gemini call. Tagging customer_id and feature is optional, and rollups are retroactive once you add the tags.
How is cached input priced?
Cache-aware. Anthropic prompt-cache reads and OpenAI cached_tokens are priced separately from fresh input; cache creation is billed as regular input. The breakdown lines up with the provider invoice.
What is cost-per-successful-run?
Total cost over the runs that finished cleanly. It is the headline metric finance can build a budget around, because it ignores money burned on runs that failed halfway.
How do spend alerts catch runaway LLM costs?
When an agent breaks its own 14-day baseline (today more than 2.5 sigma above the trailing mean), the agent's alert route fires. You find the runaway before the invoice does.
Can Agent Ping stop a runaway agent, not just tell me about it?
Yes, if you opt in. The spend guard is one call at the top of a scheduled script that checks your rules against real cumulative spend before the run starts. Once a rule trips, or you hit pause in the dashboard, the next run refuses to start. Team and Business plans.

Point Spend at your agents. See it in minutes.

Spend tells you what every agent really costs. Two lines of code, or one curl. Live in minutes, free to start, 14-day trial on paid plans.

Start free See everything Agent Ping does

Free to start. No card. The SDK never blocks your agents.