Free
$0
free forever, no expiry
50,000 spans / month — about 3,500 agent runs. No card.
- 7-day trace retention
- Full trace explorer
- 2 alert rules
- 1 seat
Pricing
Per span.
annual saves 20%
Per span.
Published overage.
One span is one unit of work: a tool call, a model call, an HTTP retry. A typical agent run emits 12–16 of them. The rate is on this page, the overage is on this page, and nothing is gated behind a call.
Team most teams
$79
per month, billed monthly
2,000,000 spans included. Overage $0.28 per 10,000 spans.
- 30-day retention
- Unlimited alert rules, PagerDuty + Slack
- Cost attribution by tenant
- 500 replays / month
- Unlimited seats
Scale
$449
per month, billed monthly
20,000,000 spans included. Overage $0.19 per 10,000 spans.
- 90-day retention, 1-year cold archive
- Unlimited replays
- EU or US data residency
- SSO / SCIM, audit log
- 99.9% ingest SLA
Enterprise
Talk to us
annual contract, invoiced
Self-hosted collector, committed volume, or both. Priced on spans, never on seats.
- In-VPC collector, Helm chart
- Custom retention up to 3 years
- DPA, SOC 2 report, pen-test summary
- Named engineer, 4-hour P1
Which plan
Drag your monthly span volume.
$79
estimated monthly bill
2.0M
Team
inside the included allowance
A span is emitted for each tool call, model call, and outbound HTTP request inside a traced run. If a run retries six times, that is six spans — which is the whole point: you are billed in proportion to how much your agents actually do.
Cost calculator
What does the fleet
What does the fleet
actually cost?
Model spend dwarfs observability spend, and that is the argument. Set your volume below — the numbers are list prices as of September 2026, not a discount you have to negotiate.
120K runs
18K tokens
Assumes a 78 / 22 split between prompt and completion tokens, and 14 spans per run — the median across accounts on Team.
Total monthly spend
$0
Prompt tokens $0
Completion tokens $0
Nocturne $0
- Cost per run, all in
- $0.0000
- Spans emitted / month
- 0
- Nocturne line item
- $0
- Observability as % of spend
One caught retry loop of the kind on our home page pays for a year of Team. That is the entire business case, and we would rather write it down than imply it.
Compare
Everything, side by side.
| Feature | Free | Team | Scale | Enterprise |
|---|---|---|---|---|
| Included spans / month | 50K | 2M | 20M | Committed |
| Overage per 10K spans | — | $0.28 | $0.19 | from $0.11 |
| Trace retention | 7 days | 30 days | 90 days | up to 3 yrs |
| Alert rules | 2 | Unlimited | Unlimited | Unlimited |
| Replays / month | 10 | 500 | Unlimited | Unlimited |
| Cost attribution by tenant | — | Yes | Yes | Yes |
| Data residency | EU | EU or US | EU or US | Your VPC |
| SSO / SCIM | — | SSO | SSO + SCIM | SSO + SCIM |
| Support | Community | Email, 1 day | Slack, 4 hrs | Named engineer |
FAQ
The questions we
The questions we
actually get asked.
Something missing? Email the engineers — there is no SDR in that inbox.
Any unit of work inside a traced run: the run itself, each model call, each tool call, and each outbound HTTP request the patched transport sees. A run that calls two tools and one model emits five spans including the root. Internal SDK retries count — they are the thing you most need to see.
Ingest keeps working and you are billed the published overage rate at the end of the cycle. Nothing is dropped silently. You can set a hard cap in project settings, in which case we start head-sampling at your ceiling and tell you in the UI and by email rather than refusing spans outright.
Not unless you turn them on. By default the SDK sends span names, timings, status codes, token counts and hashed tool arguments. Bodies are opt-in per service, and the redaction pass runs client-side before export, so redacted fields never reach our network at all.
The wire format is OTLP and you can walk away with your data any time — that part is deliberate. What we add on top is agent-shaped: loop detection over repeated tool calls, token and dollar attribution rolled up per tenant, and replay that re-executes a stored trace with recorded tool responses. A generic APM tool will show you the spans and stop there.
Yes, on Enterprise. The collector and span store ship as a Helm chart and a Docker Compose file; the UI runs against your instance. Rules, replay and cost attribution all work offline. What you give up is our on-call rotation watching the ingest path.
No. Seats are unlimited on every paid plan. Charging your product manager to look at a trace is a good way to make sure nobody looks at traces.
50,000 spans, free, forever.
Instrument one agent this afternoon. If it never loops, you never pay us a cent.