Companies now build AI into their own software: a support bot that drafts replies, product search, a pipeline that reads contracts and invoices, an agent that triages tickets, a nightly job that summarizes the day's activity, a CI step that reviews every pull request. Each of these is code the company wrote that calls model APIs in production, separate from the AI tools its employees use, such as Claude Code. To run these applications as a business, a company needs their spend attributed by application, environment, feature and customer, as well as by model, vendor and region. That attribution gives the cost of serving each customer and each feature, the numbers to price an AI feature on, and the warning when a runaway job starts spending before the invoice arrives.
Attributing Spend With Multi-Router
Each application connects to Multi-Router through its own endpoint and its own credential, so every request is attributed to the application that sent it. The support bot, the contract pipeline and the CI reviewer are separate lines of spend from the first request, and a production and a staging deployment of the same application can be separate lines too.
Each request is priced the moment it completes, from the exact model, vendor, variant and region that served it, at the price in effect at that time. Input, output, cache read and cache write tokens are each priced at their own rate, so the cost matches what the vendor charges, and the savings from prompt caching show up per application.
Inside an application, each request can also say what it was for:
| Label | Sent as | Example |
|---|---|---|
| Feature | The X-MR-Feature-Id header | support/triage, support/reply |
| Customer | The user field on Chat Completions, or metadata.user_id on Messages | tenant_acme |
| Session | The X-MR-Session-Id header | One conversation or one job run |
| Correlation | The X-MR-Correlation-Id header | A ticket or request id from the company's own logs |
With these labels, spend breaks down from the application to the feature, to the customer, down to a single conversation, and any request ties back to the company's own logs.
Overview · Traffic
| Time | Request ID | Router | Alias | Model | Customer | Feature | In/Out | Cost | Status |
|---|---|---|---|---|---|---|---|---|---|
| 14:32:11.902 | req_2Vd6Lp8NsR | rt_5Kp8Xv2QmD | support-bot | anthropic/claude-haiku-4.5 | tenant_globex | support/triage | 8,420/214 | $0.0035 | 200 |
| 14:32:09.640 | req_9Bt3Hw5KyE | rt_5Kp8Xv2QmD | support-bot | anthropic/claude-sonnet-5 | tenant_initech | support/reply | 31,250/1,406 | $0.03 | 200 |
| 14:32:07.418 | req_7Gm4Qx9WcT | rt_5Kp8Xv2QmD | support-bot | anthropic/claude-sonnet-5 | tenant_acme | support/triage | 46,000/980 | $0.03 | 200 |
req_7Gm4Qx9WcT200anthropic/claude-sonnet-5bedrock Full Request Payload → Latency 4,182ms Cost $0.03 Tokens 46,000 → 980 TTFT 1,104ms Attempts 1 Router overhead 1.4 ms Request
Routing
Prompt binding
Latency
Tokens
Cost
Provider response Request upload 0.2 ms Router · overhead 1.4 ms Upstream200 4,180.0 ms Router · response 0.3 ms Total 4,181.8 ms | |||||||||
| 14:32:04.077 | req_4Rc8Mj2FaZ | rt_5Kp8Xv2QmD | support-bot | anthropic/claude-haiku-4.5 | tenant_acme | support/triage | 6,180/342 | $0.0037 | 200 |
| 14:31:58.315 | req_6Yn1Ts7GbP | rt_5Kp8Xv2QmD | support-bot | anthropic/claude-sonnet-5 | tenant_globex | support/reply | 0/0 | $0 | 429 |
Workload Identity, No Keys
Production services and CI jobs can call Multi-Router with the identity their platform already gives them, in place
of a stored API key: a Kubernetes service account on Amazon EKS, Google GKE or Azure AKS, or a GitHub Actions,
GitLab or CircleCI job. There is no key to leak or rotate, and Traffic records the exact workload behind each
request, such as the payments/api service account or the billing repository's production deploy, so
spend is attributed to the service or pipeline that made the call.
LLM API Keys in CI and Production Workloads Considered Harmful
covers workload identity in detail.
- Request ID
- req_3Vk8Np2HsQ
- Timestamp
- 2026-11-12 09:14:52.087
- Organization
- org_3Fh8Wq2LdK
- API key
- —
- Identity type
- CI
- Machine issuer
- mii_6Gd2Tz8QwM
- Machine subject
- repo:example-co/billing:environment:production
- Machine binding
- mib_K4vB9mZp3L
- Correlation ID
- —
- Customer
- —
- Session
- —
- Peer IP
- 198.51.100.37
- Alias
- —
- Router ID
- rt_8Mc5Jx3VbN
- Model requested
- —
- Model served
- anthropic/claude-sonnet-5
- Vendor
- anthropic
- Variant
- standard
- Region
- global
- Wire id
- claude-sonnet-5
- Prompt ID
- —
- File path
- —
- Revision
- —
Cost Per Customer and Per Feature
Traffic turns those labels into the numbers a product team needs. It filters, charts and breaks down spend by
feature and by customer, the Customers view ranks customers by usage and cost, and stats such as cost per request and cost per 1,000 tokens can be added to any view. Spend per feature or per customer also
breaks down by model and vendor, so a team sees which model drives the cost of each feature and what moving it to
another model would save. For example, support/triage might cost $1,840 this month, with $1,210 on
Claude Sonnet 5 and $630 on Claude Haiku 4.5. A label prefix groups related features, so all of
support/ can be compared feature by feature.
That gives a company its unit economics: the cost of serving each customer, the cost of each AI feature, the margin on each plan, and the customer or feature behind a jump in spend. Saved views keep the reports a team opens every month, Traffic exports to CSV, and the Traffic API returns the same requests, stats and series for the company's own billing system or data warehouse.
Customers · cost by feature · Traffic
| Time | Request ID | Router | Model | Customer | Feature | Cost | Latency | Status |
|---|---|---|---|---|---|---|---|---|
| 14:32:11.482 | req_Lw4Tn8Vb2k | support-bot | anthropic/claude-sonnet-5 | tenant_acme | support/triage | $0.01 | 1,942.6 ms | 200 |
| 14:32:09.107 | req_9Pc1RmYe4s | support-bot | anthropic/claude-haiku-4.5 | tenant_globex | support/reply | $0.0019 | 641.3 ms | 200 |
| 14:32:07.845 | req_8Fq2LmXo1c | support-bot | anthropic/claude-haiku-4.5 | tenant_initech | support/summarize | $0.0031 | 808.4 ms | 200 |
| 14:32:04.219 | req_3Zr6KdWq8n | support-bot | anthropic/claude-sonnet-5 | tenant_acme | support/triage | $0.03 | 3,214.9 ms | 200 |
| 14:32:02.660 | req_7Hb5NsTq3w | support-bot | anthropic/claude-haiku-4.5 | tenant_hooli | support/reply | $0.0014 | 412.0 ms | 200 |
| 14:32:00.093 | req_2Kd8VxRe6m | support-bot | anthropic/claude-sonnet-5 | tenant_umbrella | support/summarize | $0.03 | 2,871.5 ms | 200 |
| 14:31:58.731 | req_6Ym2HcPa5t | support-bot | anthropic/claude-sonnet-5 | tenant_globex | support/reply | $0 | 212.4 ms | 429 |
| 14:31:55.318 | req_5Wt9GpLc1z | support-bot | anthropic/claude-sonnet-5 | tenant_acme | support/triage | $0.02 | 2,486.1 ms | 200 |
| 14:31:52.904 | req_4Qn3BvYk7r | support-bot | anthropic/claude-haiku-4.5 | tenant_initech | support/reply | $0.0022 | 690.7 ms | 200 |
| 14:31:49.556 | req_8Jv1MzDs2p | support-bot | anthropic/claude-sonnet-5 | tenant_globex | support/triage | $0.02 | 2,103.8 ms | 200 |
Fine-Grained Budgets for Applications
A budget can cover one application, a chosen set of applications, or every application with a tag such as
production, which includes new services as soon as they carry the tag. Each budget holds a total, a
limit for each application inside it and limits for each model. A team can cap its production support bots at $5,000
a month in total, keep any one application under $2,000, and hold its most expensive model under $1,000.
Rules act before the bill does. A rule can alert when spend is projected to reach a limit within the next two hours, post to the on-call webhook, and block only the application that crossed its line while every other application keeps serving. Budgets stack, so a team budget, an application's budget and a model limit all apply together, and the first one reached takes effect.
All budgets
| Name | Scope | Period | Spend / effective limit | Used | Status | Updated | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Production API bdg_7Qm2Xc9LpR | Tag · production | Monthly | $4,310/$5,000 | 86% | Nearing | 2h ago | ||||||||||||||
Production APIbdg_7Qm2Xc9LpRTag · productionMonthly Detailed usage → Spent $4,310 Effective limit $5,000 Remaining $690 Per-router limit $2,000base Period Sep 1 – Sep 30resets Oct 1 Pending requests 0 Routers over cap 0 of 4 Spend per day · daily · Sep 1 – Sep 30 SpendEffective limit/day $0$100$200$300$400 Effective $166.67/day Sep 1 Sep 8 Sep 15 Sep 22 Sep 29 Request approvers Requests onEmailed when someone asks for a higher limit in the Portal. The Portal hides its request button while this list is empty. priya@example.com name@example.com Add1 of 10 Budget Overrides · this period Sep 1 – Sep 30 · UTC. Each override replaces one limit until the period ends.
Pending requests Approving creates a Router override for this period. Adjust the amount before approving. No pending requests. | ||||||||||||||||||||
Standard Engineering · claude-codePolicy1 request bdg_2Lk5Pz8NcV | Tag · policy:standard-engineering | Monthly | $288/$200 per router | — | No cap | 1d ago | ||||||||||||||
Support routers bdg_9Bv3Rt6YfJ | Routers · 3 | Monthly | $1,500/$1,500 | 100% | Capped | 3d ago | ||||||||||||||
Engineering org bdg_4Hn8Ws1KdT | Org | Monthly | $11,842/$20,000 | 59% | On track | 6d ago | ||||||||||||||
Research sandbox bdg_6Tx1Jm4GqW | Routers · 1 | Lifetime | $640/$2,000 | 32% | Lifetime | 12d ago | ||||||||||||||
People, Tools and Applications
A company's AI spend comes from two places: the people who use AI tools at work, and the applications the company builds. Attributing AI Cost to People and Tools covers the first, and this post the second. Multi-Router attributes and budgets both in one place, priced the same way, so the company sees all of its AI spend in one set of numbers. The Observability and Budgets & Spend Control docs cover each view and setting in detail.