See what you could cut from your AI bill. Free, and ready in about fifteen minutes.

Connect an OpenAI or Anthropic admin key or an xAI (Grok) team management key; we only call endpoints that read cost and usage. We import 30 days of cost data, run six deterministic detectors and give you a report: what you spent, where, what looks wrong, and what we couldn’t tell from billing data alone.

Start free audit See a sample report

No card. Revoke the key at the provider any time.

AuditsAgent Waste Audit, 27 Sep – 26 Oct 2026Synthetic example

Progress

Completed in 8 min 41 s

Key validated0:00
OpenAI organization admin key. Cost and usage endpoints responded; stored encrypted.
Queued0:02
Free audit, 30 complete UTC days, 1 connection.
Syncing0:03
conn_openai: 1,284 cost records (1,284 new, 0 revised, 0 superseded)
Analyzing8:37
Building snapshot and running detectors.
Completed8:41
3 new findings. Report and export ready.

Summary

Period: 27 Sep – 26 Oct 2026
Coverage: 30 of 30 days
Final days: 27 · 3 provisional
Observed spend: $24,660.74
Findings: 3
Confirmed savings: $0.00
Snapshot: sha256:9f2c41d8…e07a

Free, no card

One 30-day audit on the free plan. Nothing to cancel.

About fifteen minutes

Two to connect, usually under ten to import, seconds to analyse.

Read access only

Cost and usage reports only. Encrypted, revocable, never shown again.

A report you can forward

In the app, and as CSV or JSON with the method behind every number.

Four steps, each one visible. You can watch them happen in the app.

Nothing runs in the background that you can’t see. Each step writes its progress to the audit, and every action is recorded in the activity log.

About 2 minutes
Connect a key (read access only)
- Paste an OpenAI organization admin key, an Anthropic Admin API key or an xAI management key. Up to two connections on the free plan.
- We call the provider’s cost and usage endpoints once before saving anything. A rejected key (401) or a key without access to usage data (403) is reported with what to fix.
- The key is encrypted with AES-256-GCM and never shown again, in the app or the API.
OpenAI
Healthy
Key type
Organization admin key
Cost API
200 OK
Usage API
200 OK
Stored
AES-256-GCM, not retrievable
Usually under 10 minutes
Import and reconcile 30 days
- Daily cost records for 30 complete UTC days, by provider, project, workspace, model and line item.
- The last 3 days are provisional and re-reconciled on every sync; revisions update in place instead of being counted twice.
- Records that disagree on a finalised day are flagged as conflicts, not silently summed.
Coverage, 30 days30 / 30
Cost records
1,284
Final days
27
Provisional days
3 (24–26 Oct)
Conflicts
0
Seconds
Run deterministic detectors
- The imported records are frozen into a snapshot and hashed. Six versioned detectors run on it; the same snapshot always gives the same findings.
- Each finding gets an evidence class, a confidence with reasons, its assumptions and the provider records behind it.
- Detectors that need inputs you haven’t given (a monthly budget, outcome telemetry) say so instead of guessing.
- spend_spike 1.0.01 finding
- spend_concentration 1.0.01 finding
- outcome_data_gap 1.0.01 finding
- sustained_growth 1.0.0No finding
- budget_burn 1.0.0No finding · no budget set
- failed_run_spend 1.0.0No finding · no telemetry
Ready when the import finishes
Read and share the report
- A report in the app with an executive summary, coverage, spend profile, findings by evidence class, next steps and methodology.
- Export it as CSV or JSON, with the methodology, coverage and every finding’s classification and assumptions.
- Invite up to three teammates on the free plan, with roles such as billing_admin or read_only.
Executive summary
$24,660.74 observed over 30 days. One anomaly worth reviewing, one concentration risk. No savings confirmed.
Anomalous spendObserved spend

Six sections, one question each. Written so finance can read it without a call.

The report shown is a synthetic example. The structure is the one your audit uses, in the app and in the export.

Agent Waste AuditNorthwind Labs, 27 Sep – 26 Oct 2026Synthetic example

Executive summary

$24,660.74 observed over 30 days. One anomaly worth reviewing, one concentration risk. No savings confirmed.

Observed spend

$24,660.74

To review

1 anomaly

Coverage

30 of 30 days

Confirmed savings

$0.00

Coverage and data quality

ConnectionDaysStatus

OpenAI · organization30 / 30Healthy

27 final days, 3 provisional (24–26 Oct), 0 conflicts, 1,284 cost records.

Spend profile

gpt-4.195.5%$23,575.48
gpt-4.1-mini3.2%$812.36
text-embedding-3-small1.1%$272.90

Findings

Spend spike on project support-agentAnomalous spend$2,779.94
95.5% of attributed spend on one modelObserved spend$23,575.48
No outcome data: cost per success unknownObserved spendUnavailable

Next steps

Check what changed on project support-agent on 15–16 Oct
Set a daily spend alert above $1,500.00
Send outcome telemetry to measure cost per successful task

Methodology

Detectors: 6 run · all 1.0.0
Snapshot: sha256:9f2c41d8b07e5a63c1d4e8f2a90b7c35e6d1f40a2b8c97e3d5f6a1c0b4e7e07a
Money: Exact integer micros; currencies never summed

Executive summary
One plain sentence about the period, and four numbers: observed spend, what to review, coverage, and confirmed savings, which an audit always reports as $0.00.
Coverage and data quality
Which days and connections the numbers rest on, how many days are final or still provisional, and any conflicts. Gaps are shown, never filled.
Spend profile
Daily spend with spike windows marked, and the largest projects, workspaces and models with their share.
Findings by evidence class
Ranked by amount. Each carries its class, its evidence and its detector, so a spike is never read as a saving.
Next steps
Concrete actions per finding: what to check, which alert to set, what telemetry would make the number knowable.
Methodology
Detector names and versions, the snapshot hash of the exact data analysed, and how money is stored and never summed across currencies.
Open the full sample report

What we can and cannot see. Stated before you connect anything.

Billing data is good at telling you what was spent and bad at telling you whether it was worth it. The report says which is which.

What we can see

Daily cost: Per provider, project, workspace, model and line item, as the provider bills it.
Usage aggregates: Token and request counts by model, from the provider’s usage reports.
Data freshness: Which days are final, which are still provisional, and when each connection last synced.
Outcomes, if you send them: Run id, status, how it was verified and a cost estimate, on paid plans, for cost per verified successful task.

What we cannot see

Prompts and completions: Connectors never read them, and telemetry rejects fields that look like content or personal data.
Whether a task succeeded: Billing data shows spend, not results. Without telemetry, we say cost per success is unknown.
Who inside a shared project: If many customers or teams share one project or workspace, the provider data can’t split them, so neither can we.
Other providers and real time: Azure OpenAI, Bedrock and Vertex aren’t supported yet. Billing data lags by hours, so nothing here is real-time.

How we handle your key and your data. Briefly, and without certifications we don’t have.

The security overview has the detail. We hold no SOC 2 or ISO 27001 certification.

Key storage: Envelope-encrypted with AES-256-GCM. Never returned by the API or shown in the app after you save it.
Key use: Read-only calls to the provider’s cost and usage report endpoints. Nothing sits in your inference path.
Revocation: Revoke the key at OpenAI, Anthropic or xAI at any time, or disconnect it in the app. Syncing stops either way.
Content: No prompts, completions or files. Your data isn’t sent to AI models or used to train them.
Retention: The free plan keeps provider cost and usage data for 90 days; paid plans keep it for 400 days or more.
Deletion: Delete your organization and its data is permanently deleted after 14 days, except records the law requires us to keep.
Isolation: Every query is scoped to your organization and covered by isolation tests.

See also the privacy policy, AI data policy and subprocessors.

After the audit. Keep the report, or keep watching.

The free plan includes one audit. If the report shows something worth watching, paid plans re-scan on a schedule and add the tools to prove a fix worked.

Free audit

USD 0

What you keep after the audit

1 audit of up to 30 days
The full report, with CSV and JSON export
Up to 2 provider connections
3 team seats, with roles
1 alert rule
REST API and MCP access
90-day data retention

Start free audit

Paid plans

From USD 299/month

Starter and Growth, billed annually

Scheduled monitoring every 24 h (Growth: every 6 h)
Up to 20 alert rules (Growth: 200), with signed webhooks
Outcome telemetry for cost per verified successful task
The verification workflow: proposals, second-person approval, experiments
Audits of up to 90 days (Growth: 180)
400-day retention (Growth: 760)

Compare plans

Taxes may apply. The pricing page has every limit and the enterprise option.

Questions about the free audit.

Which key do I need?

An organization admin key from OpenAI, an Admin API key from Anthropic, or a management key from xAI. Ordinary project or API keys can’t read organization cost reports, and the validation step will tell you if the key lacks access.

Will it slow down or change our agents?

No. Audon reads billing reports out of band. It doesn’t proxy traffic, change routing or touch your prompts.

Why does the report say “no savings confirmed”?

Because an audit can’t confirm one. It finds anomalous spend and observed patterns, and sometimes estimates a potential saving. A saving is verified only after an approved experiment measures lower cost per successful task, on a paid plan.

What if our spend is spread across many projects?

The report breaks spend down by provider, project, workspace and model. Spend the provider can’t attribute is shown as unattributed rather than guessed.

Can I run it again next month?

The free plan includes one audit. Paid plans run audits on a schedule and on demand. See pricing.

Can our own agent run the audit?

Yes. The REST API and MCP server can start an audit, read findings and export the report with scoped OAuth. See the MCP setup.

Thirty days of agent spend, audited. Before your next invoice.

Start free audit See a sample report

Free. Read access only, no card, revoke any time.