Same results. Smaller AI bill.

Find the slack in your OpenAI, Anthropic and xAI spend, and prove every saving on real traffic before it counts.

Start free audit See a sample report

Find out what you could save. Free 30-day audit, read access only, no card.

OverviewNorthwind LabsSynthetic example

Observed spend: $24,660.74; 27 Sep – 26 Oct
Potential saving: $1,140.06; Largest of 3 open levers
Verified savings: $3,420.18; 2 approved experiments
To review: 1 anomaly; $2,779.94 above median

Daily spendObservedSpike

Savings levers30 days

Prompt cachingVerified$1,271.78
Batch processingVerified$2,148.40
Model routingExperiment$1,140.06
Output lengthPotential$498.20
Retries and failed runsNeeds telemetry—

Spend by projectObserved, 30 days

support-agent$13,284.12
eval-ci$5,912.40
research-agent$3,871.53
triage-bot$1,592.69

Next steps3 open

Review the spike on support-agent15–16 Oct · $2,779.94 above median
Model routing experiment on support-agentRunning on a production canary
Send outcome telemetry for triage-botSizes retries and failed runs

Works with all common agent stacks

Agents SDK (OpenAI Agents SDK)
Agent SDK (Claude Agent SDK)
LangChain
CrewAI (CrewAI)
Microsoft AutoGen
AI SDK (Vercel AI SDK)
Dify

Read access only

Validated, encrypted, revocable. Nothing sits in your inference path.

No prompts, no outputs

Connectors read cost and usage reports only.

OpenAI, Anthropic and xAI

Provider cost and usage APIs, reconciled daily. xAI (Grok) is new.

REST and MCP

Every capability in the app is available to your agents, with scoped OAuth.

A token chart shows what you spent. Audon shows what it bought, what looks wrong, and what a change actually saved.

Find spend that doesn’t fit

Deterministic detectors compare every project, workspace and model against its own history: spikes, concentration, sustained growth, budget burn, failed runs.

See the evidence behind every number

Each finding links to the provider records it came from, its baseline, its confidence and what could make it wrong.

Count savings only when they’re measured

Fixes go through approval by a second person and a controlled experiment on production traffic. Only then does an amount become a verified saving.

From audit to savings. Spend less for the same results, and prove it.

The audit shows where the money goes. The savings plan sizes each lever on your own data, your team makes the change, and an experiment on production traffic shows what it actually saved.

SpendLast 30 daysSynthetic example

27 Sep – 26 Oct

Observed spend

$31,482.16

Anthropic

$18,906.42

60.1%

OpenAI

$10,733.18

34.1%

xAI

$1,842.56

5.9%

Uncached input$11,963.22
Cached input$1,888.93
Output$16,055.90
Batch$1,574.11

Project · modelSpend

support-agent claude-sonnet-5-5$12,640.20
catalog-enrich gpt-5$7,912.64
research-agent claude-opus-5-5$6,266.22
catalog-enrich gpt-5-mini$2,820.54
triage-bot grok-4$1,842.56

See
Every dollar by provider, project, model and token type, reconciled daily.
SpendLast 30 daysSynthetic example
27 Sep – 26 Oct
Observed spend
$31,482.16
Anthropic
$18,906.42
60.1%
OpenAI
$10,733.18
34.1%
xAI
$1,842.56
5.9%
Uncached input$11,963.22
Cached input$1,888.93
Output$16,055.90
Batch$1,574.11
Project · modelSpend
support-agent claude-sonnet-5-5$12,640.20
catalog-enrich gpt-5$7,912.64
research-agent claude-opus-5-5$6,266.22
catalog-enrich gpt-5-mini$2,820.54
triage-bot grok-4$1,842.56
Find
Every finding with its evidence, its baseline and what happened next.
FindingsSpend spike on project support-agentSynthetic example
FND-0142High
Spend spike on project support-agent
Two days at 3.4× the 14-day median, right after a deploy.
Classification
Anomalous spend
Above median
$2,779.94
Period
15–16 Oct 2026 (UTC)
Confidence
High
Status
Experiment running
Daily spend · support-agentSpike window
Detected by scheduled audit2 days ago
Note from Priya Raman1 day ago
Retry storm after release 4.12.
Approved by a second person20 h ago
Cap retries at 2. Rollback: max_retries=5.
Experiment running on production canarynow
Size
A potential saving per lever, with its formula and one assumption you can change.
Savings planPrompt cachingSynthetic example
Prompt caching
support-agent · claude-sonnet-5-5
Potential saving
Potential saving, 30 days
$806.00 potential saving at 25% of uncached input moved to cache
25% × $4,960.00 × (1 − 0.35)
Moved to cache25%
Stated assumption: 25%.
Assumptions
Uncached input
$4,960.00
Prefix reuse
At least 4×
Cached cost
0.35× base input
Cache reads 0.05× and 5-minute writes 1.25× base input on Sonnet 5.5. Anthropic documentation, checked 26 Oct 2026.
Prove
Approval by a second person, then an experiment on production traffic.
OptimizationsOPT-0031Synthetic example
Cache the system prompt on support-agent
OPT-0031
Verified
Proposal from the lever · 28 Sep
Rollback: remove the cache breakpoint.
Approved by a second person · 29 Sep
Experiment on production traffic · 14 days
Control214
Treatment188
Successes, 30 needed per window · no quality or latency regression
Verified by measurement · 13 Oct
Verified saving, 14 days
$642.18
Cost per success $0.142 → $0.118
Keep
An alert on the changed project, so the saving doesn’t quietly erode.
Alertssupport-agentSynthetic example
Cost per success above $0.13 on support-agent
cost_per_success > $0.13
Armed
Cost per successful taskAlert at $0.13
Verified saving
$642.18
Alerts fired
0
Next check
Next full day

Five savings levers, sized on your own data.

Each lever applies published provider rates and one stated assumption. Move the assumption and the number moves with it.

Potential, 30 days
$806.00
25%
of uncached input moved to cache
Prompt caching
Uncached input per model, from token-type line items. Cache reads cost 0.1× input on Anthropic, at most half on OpenAI, 75–85% less on Grok 4.x.
Potential, 30 days
$1,073.32
20%
of spend that can wait 24 h
Batch processing
Standard-tier spend that could wait 24 h. Batch costs 50% less on OpenAI and Anthropic, and 20% less on Grok 4.x.
Potential, 30 days
$1,384.71
20%
of premium-model tokens routed
Model routing
Your blended cost per token for each model, so no price list is needed. Per-token rates differ 10–20× between frontier and small models.
Potential, 30 days
$639.15
15%
fewer output tokens
Output length
The output share of each model’s cost, once it reaches 60%. On most models, output tokens cost 4–5× input.
Potential, 30 days
$212.40
41 verified failed runs, no assumption
Retries and failed runs
Failed runs and retries from the outcome telemetry you send. No provider rate involved: it is what failed runs already spent.
How we size a lever
Versioned rules on your own cost data, with one stated assumption you can change. Levers can overlap, so they are never added up.
Read the method

Amounts are a synthetic example. Provider facts checked 26 Oct 2026. A lever writes a potential saving; only an approved experiment makes it verified.

Providers we read

Read-only keys on cost and usage reports. No prompts, no outputs.

OpenAI (OpenAI)
Costs and Usage APIs, with cached tokens, the batch flag and service tier.
Cost Report by token type and service tier, plus the usage report.
xAI (xAI)New
Management API billing usage, per team and model.

An audit you can forward to finance. Ready in minutes, on 30 days of your real spend.

Connect a read-only key and get a report with a summary, coverage, findings with their evidence, and the method behind every number.

How the free audit works Open the sample report

AuditsNorthwind Labs, 27 Sep – 26 OctSynthetic example

Executive summary

$24,660.74 observed over 30 days. One anomaly to review. $3,420.18 verified by two approved experiments, not by this audit.

Observed spend: $24,660.74
To review: 1 anomaly
Coverage: 30 of 30 days
Verified savings: $3,420.18

Spend spike on project support-agent$2,779.94
95.5% of attributed spend on one model$23,575.48
Cost per success unknownNo outcome data

AlertsAll rulesSynthetic example

RuleStatus14 days

Daily spend above $1,500

daily_spend > $1,500.00

Triggered 16 Oct

Monthly budget 80% reached

budget_burn ≥ 80%

Armed

Provider sync stale for 12 h

sync_staleness > 12 h

Armed

Cost per success above $0.13

cost_per_success > $0.13

Armed

Catch the next spike while it’s small. Not on next month’s invoice.

Paid plans re-scan on a schedule and evaluate alerts on complete days, so a runaway retry loop shows up in a day, not in a month-end review.

Monitoring

Daily spend and budget-burn alerts
Per-project and per-model thresholds
Stale-sync and degraded-connector alerts
Email, signed webhooks or your agent
Dry-run any rule against history
Acknowledge and track every event

Your agents can run it. Within the limits a person sets.

Every capability in the app, all 91 of them, is available through a versioned REST API and a remote MCP server with scoped OAuth. Decisions that move money or accept terms still need a signed-in person.

For developers

Remote MCP server, OAuth 2.1 + PKCE
Versioned REST API and OpenAPI spec
Scoped API keys, intersected with roles
Idempotent writes, signed webhooks
Outcome telemetry for cost per success
Human approval where it matters

MCP setup REST API

DevelopersMCP session

› Audit our AI spend for October and set a daily alert.

  audit_start      { period_days: 30 }       → queued
  findings_list    { sort: "impact" }        → 3 findings
    · spend spike, project support-agent
      amount_is_verified_saving: false
  alert_create     { metric: "daily_spend" } → armed

  optimization_approve { … }
    ✕ approval_required: a person approves this in the app

Built for teams whose margin runs through agents. See what it finds in yours.

AI-native SaaS

Gross margin depends on what every customer’s agents cost you.

Read how it works for you

Customer support

Resolution agents that retry, loop or escalate quietly cost the most.

Read how it works for you

Financial services

Every AI cost decision needs an approval trail an auditor can follow.

Read how it works for you

Agencies & consultancies

Know which client’s agents eat the margin before the invoice goes out.

Read how it works for you

E-commerce & marketplaces

Catalogue, search and shopping agents scale with traffic, and so do their costs.

Read how it works for you

Not on the list?

If you pay OpenAI, Anthropic or xAI for agents, the free audit works the same way.

Start with the free audit

Built to pass your security review. Least data by design, every control documented, nothing claimed that isn’t true.

Request the security packet Security overview

Read access only
We call cost and usage endpoints. Nothing sits in your inference path.
Keys envelope-encrypted
AES-256-GCM with a separate master key. Never shown again, destroyed on disconnect.
No prompts, no model training
We never read prompts or outputs, never send your data to an AI model, never train on it.
Tenant isolation, tested
Every record is scoped to one organization. Automated tests cover web, REST and MCP.
Tamper-evident audit log
Every sensitive action, by a person or an agent, in a hash-chained log you can verify.
Human approval
Roles, scoped agent access, and four-eyes approval before any change counts.

Questions teams ask before connecting.

What access do you need?

An organization admin key for OpenAI, an Admin API key for Anthropic or a management key for xAI. We use it only to read cost and usage reports. It is encrypted with AES-256-GCM, never shown again, and you can revoke it at the provider at any time.

How long does the free audit take?

About ten minutes to connect and import 30 days, then the detectors run in seconds. You get the full report in the app, plus an export.

Can it see what our agents say?

No. Connectors read billing and usage aggregates only. If you want cost per successful task, you can send outcome telemetry (run ids, status, cost estimates). The API rejects fields that look like prompts or personal data.

How do you decide something is a saving?

We don’t, on our own. A potential saving becomes validated when a person records the evidence, and verified only after an approved experiment shows lower cost per successful task without quality or latency regression, with at least 30 successes in each window.

What does it cost after the audit?

Paid plans add scheduled monitoring, alerts, the verification workflow and more connections. See pricing.

How do you help us spend less?

The savings plan sizes five levers on your own cost data: prompt caching, batch processing, model routing, output length, and retries and failed runs. Each shows its formula, its assumption and the provider facts it uses. Your team makes the change; we track the approval and the experiment, and only a measured result counts as a verified saving.

Do you support Azure, Bedrock or Vertex?

Not yet. It works when your agents call the OpenAI, Anthropic or xAI (Grok) APIs directly with your own keys, whatever framework they use; the xAI connector is new. Billing through Azure OpenAI, Amazon Bedrock or Google Vertex isn’t covered yet.

Same results. A smaller AI bill, proven.

Start free audit See a sample report

Free 30-day audit. Read access only, no card, revoke any time.

Same results. Smaller AI bill.

A token chart shows what you spent. Audon shows what it bought, what looks wrong, and what a change actually saved.

Find spend that doesn’t fit

See the evidence behind every number

Count savings only when they’re measured

From audit to savings. Spend less for the same results, and prove it.

See

Find

Size

Prove

Keep

Five savings levers, sized on your own data.

Prompt caching25% of uncached input moved to cache × (1 − effective cached cost), with each cached prefix reused at least four times.

Batch processing20% of the project’s spend can wait up to 24 h × the provider’s batch discount.

Model routing20% of premium-model tokens routed to the cheaper model × (1 − cheap rate ÷ premium rate).

Output length15% fewer output tokens (max tokens, structured output, terser formats) × output cost.

Retries and failed runsEstimated cost of verified failed runs.

Providers we read

An audit you can forward to finance. Ready in minutes, on 30 days of your real spend.

Catch the next spike while it’s small. Not on next month’s invoice.

Your agents can run it. Within the limits a person sets.

Built for teams whose margin runs through agents. See what it finds in yours.

Built to pass your security review. Least data by design, every control documented, nothing claimed that isn’t true.

Questions teams ask before connecting.

Same results. A smaller AI bill, proven.

Prompt caching

Batch processing

Model routing

Output length

Retries and failed runs