What is measured

Every figure in this section is built from one record: the model call, sent by your gateway for every call it makes. Nothing is sampled.

One model call

Field	What it is
input_tokens	Input tokens not read from the cache
cache_read_tokens	Input tokens read from the prompt cache
cache_write_tokens, cache_write_1h_tokens	Input tokens written to the cache; the second is the one-hour share where the provider reports it
output_tokens	Output tokens, including any reasoning tokens the provider bills as output
cost_usd	What the call cost, on your deployment’s price card
estimated	true where the cost was worked out from the price card rather than reported with a price by the provider
model, backend	The model, and the platform that served it (for example Bedrock in the EU)
pack, task, agent, run_id_hash	Which agent, task and run the call belongs to
latency_ms, paced_ms, throttled	How long it took, how long it waited for a rate limit, and whether the provider throttled it
outcome, reason	ok, refused or error, with a fixed reason code

All input is the sum of the first three: uncached input, cache reads and cache writes. The same convention runs through every page and every API read.

No prompt, answer or document text is in the record. See the telemetry contract.

Reasoning tokens

A thinking model’s reasoning tokens are billed as output, and are counted in output_tokens and its cost. They are measured separately by the Token Value Benchmark, which reports input, output, cached and reasoning tokens per task as four figures. A separate reasoning count on every model call: Coming

Currency

Model costs are recorded in US dollars today. Every read also carries each money figure in your display currency (*_money), converted at the European Central Bank’s rate for that amount’s own day, and never adds two currencies’ raw amounts together. Set the display currency in Settings → Preferences for yourself, or for the whole tenant in Settings → Organisation.

The benchmark records each result in the provider’s own currency (currency and cost_native) beside its dollar figure. Carrying a provider’s own currency through on every model call: Coming

Grouped the way you need it

Grouping	Read
Per call	Each model_call record, in your tenant’s export
Per run	GET /api/v1/runs/{run_id_hash} and its run map: tokens and cost of every step
Per agent	GET /api/v1/billing/usage?group_by=agent, and the agent’s own page
Per pack	GET /api/v1/usage?group_by=pack. A pack is a department’s set of agents, which is how most clients see spend by team
Per model, task or day	GET /api/v1/usage?group_by=model (or task, or day)
Per connection	GET /api/v1/billing/usage?group_by=connection: each provider connection, direct or through Atlantic
Caching, by model, agent, pack, deployment or provider	GET /api/v1/caching

curl

# Tokens and cost by pack for October
curl "https://$AAL_TENANT.atlanticagenticlabs.com/api/v1/usage?since=2026-10-01&until=2026-10-31&group_by=pack" \
  -H "Authorization: Bearer $AAL_KEY" \
  -H "CF-Access-Client-Id: $CF_ACCESS_CLIENT_ID" \
  -H "CF-Access-Client-Secret: $CF_ACCESS_CLIENT_SECRET"

Python

import os

import httpx

headers = {
    "Authorization": f"Bearer {os.environ['AAL_KEY']}",
    "CF-Access-Client-Id": os.environ["CF_ACCESS_CLIENT_ID"],
    "CF-Access-Client-Secret": os.environ["CF_ACCESS_CLIENT_SECRET"],
}
r = httpx.get(
    f"https://{os.environ['AAL_TENANT']}.atlanticagenticlabs.com/api/v1/usage",
    params={"since": "2026-10-01", "until": "2026-10-31", "group_by": "pack"},
    headers=headers,
)
r.raise_for_status()
print(r.json())

TypeScript

const q = new URLSearchParams({ since: "2026-10-01", until: "2026-10-31", group_by: "pack" });
const res = await fetch(`https://${process.env.AAL_TENANT}.atlanticagenticlabs.com/api/v1/usage?${q}`, {
  headers: {
    Authorization: `Bearer ${process.env.AAL_KEY}`,
    "CF-Access-Client-Id": process.env.CF_ACCESS_CLIENT_ID!,
    "CF-Access-Client-Secret": process.env.CF_ACCESS_CLIENT_SECRET!,
  },
});
console.log(await res.json());

Each group carries calls, input_tokens, cache_read_tokens, cache_write_tokens, output_tokens, cost_usd, estimated_calls, refused and errors. See the Usage reference.

Spend is not broken out per person: a model call carries no person, and a person appears elsewhere in telemetry only as a keyed hash. Per-person limits do exist; see Limits and alerts.

Checked against the provider’s bill

The gateway’s figure is what your software reported. Where LightHouse can read the provider’s own figure (AWS Cost Explorer for Bedrock, Anthropic’s cost report for the Claude API), Spend → Reconciliation shows both per day and connection, and the difference, without smoothing it. GET /api/v1/billing/reconciliation?month=YYYY-MM.