Cost attribution
Break spend down by model, route, project, endpoint, or your own tags — including per-customer, if you send a customer ID.
Product
Every call, token, and dollar — in real time.
Vantafold records a full trace for every request: the route taken, each tool call, the tokens consumed, and what it cost. No sampling, no separate agent to install.
Requests
4.82M
vs. last week +12.4%
p95 latency
940 ms
vs. last week −18.1%
Tokens
1.94B
vs. last week +9.7%
Spend
$12,480
vs. last week +4.2%
Requests per hour
last 24 hours · UTC
00:0006:0012:0018:0023:00
Spend by model
last 24 hours
Trace tr_5b1d9e0c
1,284 ms total
POST /v1/reason
1284 ms
gateway.auth
8 ms
router.select
31 ms
model.frontier-lg
1180 ms
tool.ledger_lookup
402 ms
observability.emit
6 ms
Interface preview. Figures are illustrative.
Answers, not just charts.
The questions teams actually ask about production AI, answerable without exporting anything to a spreadsheet.
Full-fidelity traces
Every span: gateway, routing decision, provider call, each tool invocation, and the retry chain when one occurs.
Latency breakdown
Time-to-first-token and total duration, split by the part of the path that spent it, so you know whether to blame the model or your tool.
Quality signals
Attach evaluation scores and user feedback to a trace and watch them move when you change a route or a prompt.
OpenTelemetry export
Ship spans and metrics to your own collector over OTLP. Vantafold spans join your existing traces rather than living in a separate tool.
Redaction
Field-level redaction before storage, or zero-retention mode where only metadata and counters are kept.
Two lines, then it is on.
from vantafold import Vantafold
from opentelemetry import trace
client = Vantafold(
otel_endpoint="https://otel.internal:4317", # your collector
otel_attributes={"service.name": "billing-agent", "deploy.env": "prod"},
)
with trace.get_tracer(__name__).start_as_current_span("reconcile"):
response = client.reason(model="auto", input=prompt, tools=[ledger_lookup])
# Spans, token counts, and cost are exported on the same trace as your own spans.
print(response.trace_id, response.usage.cost_usd)Trace schema
Every trace carries the same envelope, so you can query it without knowing which model answered.
{
"trace_id": "tr_5b1d9e0c",
"duration_ms": 1284,
"route": { "requested": "auto", "selected": "frontier-lg", "attempts": 1 },
"usage": { "input_tokens": 4210, "output_tokens": 186, "cost_usd": 0.0154 },
"spans": [
{ "name": "gateway.auth", "start_ms": 4, "duration_ms": 8 },
{ "name": "router.select", "start_ms": 12, "duration_ms": 31 },
{ "name": "model.frontier-lg", "start_ms": 46, "duration_ms": 1180 },
{ "name": "tool.ledger_lookup", "start_ms": 214, "duration_ms": 402 }
],
"tags": { "project": "billing-agent", "customer_id": "cus_2f81" }
}Know what your AI costs before finance does.
Traces are on by default on every plan, including the free tier.