What your agents cost

Agent Spend Optimization. Smaller AI bills, same results.

Name: Agent Spend Optimization
Brand: Audon

Agent Spend Optimization is Audon’s product for cutting what AI agents and coding tools cost. It reads OpenAI, Anthropic, xAI, Cursor and Devin billing with read-only keys, finds savings such as prompt caching, batch processing, model routing and unused seats, and counts a saving only after it is measured. Formerly Agent Loss Prevention.

Find out in minutes how much of your AI bill you could cut without changing what your agents deliver, get the exact fix for each, and prove every saving before it counts.

Start auditFree See a sample report

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: To review
Above median: $5,559.88
Period: 15–16 Oct 2026 (UTC)
Confidence: High
Status: Fix ready

Cap retries at 2

max_retries: 5 → 2

Sends a GitHub issue to northwind/support-agent

Daily spend · support-agentSpike window

Detected by scheduled audit2 days ago
Note from Priya Raman1 day ago
Retry storm after release 4.12.
Fix suggested20 h ago
Cap retries at 2. Rollback: max_retries=5.
Ready to apply, on a 10% canarynow

When you apply it

Your coding agent gets a GitHub issue

It opens a pull request behind a flag

How it’s measured

10% canary for 7 days

Cost per successful task, before and after

Rollback

max_retries: 2 → 5

One click sends the rollback

Six steps, from a key to a measured saving.

1
Connect
A read-only key per provider, encrypted. No SDK, no proxy.
2
Reconcile
Daily cost records; recent days stay provisional.
3
Detect
Versioned detectors compare each project and model with its history.
4
Explain
Every finding carries its evidence, baseline and confidence.
5
Monitor
Scheduled re-scans, with alerts by email, webhook or agent.
6
Measure
Live changes are measured from billing; an experiment verifies them.

Detectors in this release. Deterministic, unit-tested and versioned.

Every finding records the detector version and the snapshot it ran on, so a result can be reproduced. No detector can write a verified saving.

FindingsDetector registry

Detector	What it finds	Writes	Needs
spend_spike1.0.0	Days where spend on a project/workspace/model exceeds 2.5× its trailing 14-day median (minimum 7 days of history).	Anomalous spend	Cost data, 7+ days of history
spend_concentration1.0.0	A single project/workspace/model accounts for at least 60% of spend in the period.	Observed spend	Cost data
budget_burn1.0.0	Month-to-date spend run-rate projects above the organization’s configured monthly budget.	Observed spend	Cost data and a monthly budget
sustained_growth1.0.0	Week-over-week spend growth above 20% for three consecutive complete weeks on one dimension.	Observed spend	Cost data, 4+ weeks
outcome_data_gap1.0.0	Material spend with no outcome telemetry — an analysis gap, never presented as waste.	Observed spend	Cost data

How we size a lever. On your own cost data, with every assumption stated.

All sixteen savings levers are detectors too: deterministic, versioned and reproducible. Each one reads a signal in your data, applies the provider facts it records and one explicit assumption, and writes a potential saving. Nothing more.

Your data, not a price list
Signals come from your cost line items: token types, service tier, observed cost per token per model.
One stated assumption
The share of eligible traffic is never claimed as known. It is an explicit, editable assumption on the lever.
Facts with a source
Each lever records the provider facts it used, with the source and the date they were checked.
Sized in order, so they add up
Levers on the same spend are sized in a fixed order, each on what the earlier ones left.

Model providers10

Changes to how your code calls a provider’s API, or to a setting on the provider account.

Service tier
Pay the fast or priority surcharge only where someone waits; run work nobody waits on at flex or batch rates.
Regional processing
Drop the regional or US-only processing surcharge where no contract requires it.
Model choice
Move to a newer, cheaper version of the same model, or route simpler requests to a cheaper model.
Tests and evals
Run tests, evals and CI on a cheaper model, a smaller sample or a nightly schedule.
Context size
Stop resending growing history and stay under long-context price thresholds.
Output and reasoning
Cap outputs and set reasoning effort per task, so you pay for the answer, not the padding.
Embeddings
Embed each document once (skip unchanged content) and run bulk embedding at batch rates.
Retries and failed runs
Stop paying for runs that fail or retry without producing a result.
Prompt caching
Serve repeated prompt prefixes from cache, and stop paying for cache writes that are never read.
Storage and hosted tools
Delete stored files, vector stores and caches nobody uses, and cap paid tool calls.
See which ones apply to you. The free audit sizes every method on your last 30 days and shows only the ones that apply, each with its fix.
Start auditFree

Coding tools and seats, contracts and commitments6

Plans, seats and defaults an admin changes in a tool’s dashboard. Contracts and commitments are sized on the numbers you enter under Settings → Contracts.

Coding tools and seats
Right plan per person
Give each person the cheapest plan that covers how they work: remove, downgrade, upgrade or move to a seat.
Coding tools and seats
Coding model mix
Default coding agents to a cheaper model or the tool’s own model pool; keep the premium model for hard tasks.
Coding tools and seats
Cursor Max mode
Keep Max mode for large, multi-file work; normal mode costs less per token for the rest.
Coding tools and seats
Overlapping tools
People paying for two tools that do the same job: keep the one they use.
Contracts and commitments
Cloud commitments and credits
Pay for the same models through AWS, Google Cloud or Azure where a commitment or credits would otherwise go unused.
Contracts and commitments
Reserved capacity
Size provisioned throughput to what you use, or move low-use capacity back to pay-as-you-go.

From finding to fix. Apply it in one click.

Each lever comes with the exact change for your provider and model. One click hands it to your coding agent as a GitHub issue or a webhook. The saving is still measured before it counts.

The fix, on every finding
Steps, code and rollback for your provider and model.
Apply fix
One click sends it to your coding agent as a GitHub issue or a webhook.
Followed to live
The pull request and the merge show up on the finding. Then the saving is measured.

FindingsPrompt caching on claude-sonnet-5-5Synthetic example

The fix · Anthropic · claude-sonnet-5-5

Cache the static prompt prefix

1.Move instructions, tools and examples to the front
2.Mark them as cacheable
3.Ship behind a flag and watch the cached share

system: [{
  type: 'text', text: STATIC_INSTRUCTIONS,
  cache_control: { type: 'ephemeral' },
}],

Sends a GitHub issue to northwind/support-agent

Issue #214 opened
Pull request #215 open
Live on merge

AlertsAll rulesSynthetic example

RuleStatus14 days

Daily spend above $3,000

daily_spend > $3,000.00

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

Monitoring and governance by default. Hard limits stay with your provider.

Provider billing APIs report with a delay and the base product isn’t in your request path, so it doesn’t stop spend as it happens. It catches it the day it shows up and governs what you change.

Monitoring: Daily re-scans and alerts by email or webhook.
Governance: Rollback plans, a second approver, an audit trail.
Hard caps: Your provider’s limits, set where Audon suggests.
Cost per team: Spend per team, with a monthly budget for each.
Renewals and prices: Renewal reminders and announced price changes.
New keys: New or unnamed API keys wait for review.

DevelopersPOST /v1/telemetry/events

{
  "events": [{
    "run_id": "run_8f2c",
    "workflow_id": "support-triage",
    "started_at": "2026-10-21T09:14:02Z",
    "result_status": "succeeded",
    "result_verification": "system_verified",
    "estimated_provider_cost_minor": 42,
    "currency": "USD"
  }]
}

Cost per successful task. The number billing data can’t give you.

Send outcome events (run id, workflow, status, how the outcome was verified, estimated cost). Audon computes cost per verified success and finds spend on failed runs. Fields that look like prompts or personal data are rejected.

Telemetry API MCP setup

Methodology and the words we use.

All terms in the glossary

Observed spend

Invoice-aligned, provider-reported cost for the selected UTC days. Sourced from the OpenAI organization Costs API and the Anthropic Cost Report API. Recent days are provisional and are re-reconciled on every sync.

Anomalous spend

Observed spend with features that indicate a possible problem (e.g. a day more than 2.5× its trailing 14-day median). It is not a saving: it may be legitimate demand.

Potential saving

A modelled reduction if a proposed change works, under stated assumptions.

Validated opportunity

A scoped, feasible change with supporting evidence and a computable expected benefit, confirmed by a person.

Verified saving

A measured reduction in cost per verified successful task in the latest recorded experiment on production traffic, approved and verified by people other than those who proposed it and recorded the results, with no success-rate drop greater than 2 percentage points, no p95 latency increase greater than 20%, and at least 30 verified successes per window:

measured saving = (cost per success before − cost per success after) × successful tasks after

Cost per verified successful task

Cost per verified successful task = Σ attributable run costs ÷ count of verified successful runs. When the denominator is zero, or cost is missing, the value is unavailable — never 0.

Cash impact

Lower usage reduces future usage-based invoices. Under prepaid commitments it frees capacity rather than reducing near-term cash.

Monitoring vs blocking

Alerts evaluate complete UTC days of provider billing data, which can lag by hours. They are monitoring, not real-time spend blocking.

Money

Amounts are stored as exact integer micros (10⁻⁶ of the billed currency). Currencies are never summed together.

Integrations. Each with what we read and what we never read.

Questions teams ask before connecting.

What is Agent Spend Optimization?

Is this the same as Agent Loss Prevention?

Yes. Agent Loss Prevention was renamed Agent Spend Optimization in November 2026. The product, your data, your plan and the API are the same, and old links lead here.

Which providers and tools does it read?

OpenAI, Anthropic (Claude Platform), Claude Enterprise, xAI, Cursor, Devin (Enterprise), Claude Code (personal plans), Codex (ChatGPT plans). Plans with no billing API, such as Claude Pro or Max, ChatGPT, GitHub Copilot or Gemini, are added under Settings → Subscriptions. See integrations.

Does it change our code, keys or AI traffic?

No. Keys are used read-only to read cost and usage reports, encrypted with AES-256-GCM and never shown again. Audon is not in your request path. Every finding comes with the exact fix; with the Auto-apply add-on, Audon hands it to your own coding agent and your team reviews and merges the change.

When does a potential saving count as saved?

When your team marks a change as done, Audon measures it from billing, as cost per unit of work before and after, so volume changes don’t count. It becomes a verified saving only after an approved experiment shows lower cost per successful task with at least 30 successes in each window.

What does it cost?

The 30-day audit is free, with no card. Starter is USD 109 a month (USD 89 billed yearly) and adds daily monitoring and verified savings. See pricing.

Last updated 2 November 2026

Smaller AI bills. Same results, proven.

Start auditFree See a sample report

Agent Spend Optimization. Smaller AI bills, same results.

Six steps, from a key to a measured saving.

Detectors in this release. Deterministic, unit-tested and versioned.

How we size a lever. On your own cost data, with every assumption stated.

Model providers10

Service tier

Regional processing

Model choice

Tests and evals

Context size

Output and reasoning

Embeddings

Retries and failed runs

Prompt caching

Storage and hosted tools

Coding tools and seats, contracts and commitments6

Right plan per person

Coding model mix

Cursor Max mode

Overlapping tools

Cloud commitments and credits

Reserved capacity

From finding to fix. Apply it in one click.

Monitoring and governance by default. Hard limits stay with your provider.

Cost per successful task. The number billing data can’t give you.

Methodology and the words we use.

Integrations. Each with what we read and what we never read.

Questions teams ask before connecting.

Smaller AI bills. Same results, proven.