If AI coding costs have grown to $150,000 a month, there’s a more important question than “Which model did we use most?” What did we ship from the roadmap we committed to with that money?
First, let’s separate where the numbers come from. The $150,000 per month figure is the scale of a Claude bill mentioned when a Product Hunt user asked Navigara a question; it is not a publicly verified Navigara customer case. What Navigara has publicly described is a product that connects Git history, Jira or Linear, and AI spend to show costs by roadmap item and work that is not linked to one. This article uses $150,000 not as a confirmed case, but as a starting point for explaining how to turn a large bill into management-accounting data.
Model-only bills hide why spending is actually high
Costs by model are purchasing records, not performance records. A table saying “$60,000 on Opus and $50,000 on Sonnet” cannot tell you whether a checkout overhaul was expensive, incident response was expensive, or money leaked into a canceled experiment.
The data provider APIs offer generally stops at this boundary. Anthropic’s Usage & Cost API can break usage down by model, workspace, API key, and more, and retrieve costs by day. OpenAI’s organization API likewise exposes usage and costs and can group them by project ID and similar dimensions. This data is useful for reconciling invoices, but it does not tell you which product goal a project moved.
| Unit of analysis | Questions it can answer | Questions it cannot answer |
|---|---|---|
| Model·tokens | Which model incurred the cost? | What did we ship? |
| User·seat | Who used the tool? | Was that use necessary work? |
| PR·commit | Which code change was it linked to? | What value did it create for customers or the roadmap? |
| Roadmap item | How much did each goal cost? | Was that goal itself the right one? |
That is why the cost unit should not be the model, but a cost object such as an initiative, epic, or feature. Navigara also says it links AI spend to initiatives, projects, issues, PRs, and commits, then classifies it as features, maintenance, fixes, and more. The key idea is not buying a particular product. It is creating traceable links between provider billing data and actual delivery records.
The answer is a ‘dollars→sessions→PRs→issues→goals’ linkage table
Roadmap cost is not calculated in one step. You need to connect a cost event to the work session where it happened, then connect the changes from that session to the issues and goals they belong to.
Start automation with work-naming conventions. Include the issue key when running an agent, name branches like ENG-2917-checkout, and leave the same key in the PR body, and you can join data across all three systems. If you operate direct API calls, separating workspaces or projects by team can create the first cost boundary. But when one session touches multiple issues or shared seat costs are mixed in, do not force allocations down to the cent; keep them as shared costs.
GitHub Copilot can also export usage metrics and PR lifecycle data scoped to organizations, enterprises, repositories, and users. However, GitHub notes that organization-level figures are attributed based on membership and can appear duplicated, and that different Copilot API resources should not be compared directly. So reconcile provider-dashboard numbers and your own ledger to the monthly billed total before starting detailed allocation.
Do not automatically label ‘not linked to the roadmap = waste.’
Urgent incident response, security patches, refactoring, and exploratory experiments can be legitimate while absent from the roadmap. On Product Hunt, Navigara’s founder also said that unlinked work is not necessarily bad, but its reason should be explainable. Treat unlinked costs not as confirmed waste, but as a queue requiring classification.
In practice, decisions become easier when you divide costs into four buckets. First is shipped roadmap work; second, approved maintenance and risk reduction; third, shared costs such as common infrastructure and seats; fourth, unclassified costs with no evidence of linkage. Have an owner review only the last bucket and split it further into canceled experiments, duplicate work, and missing tickets.
A cost table needs confidence and quality alongside dollars
If you create only one dollar total per roadmap item, you have built another vanity metric. More PRs or commits do not mean value grew in the same proportion. Navigara proposes its own unit, ETV, as a denominator by assessing commits with LLMs and rules, but it is a proprietary metric defined by the company. Its public research report also states that its analysis is limited to merged commits in public repositories, cannot observe activities such as code review, mentoring, and incident response, and does not estimate a causal ratio between AI adoption and outcomes.
So do not treat ETV or an in-house score as truth in a single number. DORA likewise recommends combining complementary indicators such as suggestion acceptance rate, model quality, trust, and review time with existing delivery metrics when evaluating AI tools. A roadmap cost table should include at least these three axes together.
For example, this is much more useful than saying “AI cost for the checkout overhaul: $20,000”: “$14,000 directly linked, $4,000 in allocated shared cost, $2,000 estimated; released two weeks ahead of plan; three post-release rework items.” When deciding whether to cut costs, you can look first at work types with high unclassified rates and heavy rework, rather than simply cutting the most expensive model.
How to build a roadmap cost table this week
1. Bring provider charges into one ledger
For Anthropic, pull the organization-level /v1/organizations/cost_report and usage reports; for OpenAI, use /v1/organization/costs and the Usage API. Keep separately billed items, such as API, Claude Enterprise, and fixed seats, in separate columns, and first verify that the month-end total matches the actual invoice.
2. Require an issue key as the shared identifier for every task
Organize the initiative→epic→issue hierarchy in Jira or Linear, and leave the issue key in at least two of the agent prompt, branch, commit, and PR. Also create types such as INC for urgent work, MAINT for maintenance, and EXP for experiments, so legitimate work outside the roadmap does not fall into the unclassified bucket.
3. Separate automatic allocation from estimated allocation
Label costs directly connected by a session ID or dedicated workspace as “confirmed”; costs inferred only from matching time, user, or repository as “estimated”; and seats and common infrastructure as “shared costs.” When one session spans several issues, split it using pre-agreed criteria such as changed files or execution time, and record the allocation rule in the ledger.
4. Ask only three questions in the monthly review
Check: “Which goal’s unit cost changed?”, “Why did unclassified costs arise?”, and “Did rework, incidents, and review time fall along with savings?” In the first month, focus on raising the direct-link rate rather than calculating precise ROI, and do not use individual rankings in performance reviews.
Built this way, $150,000 per month changes from an intimidating total into a decision-ready portfolio. More important than the fact that you used an expensive model, you can start to see which roadmap items deserve more investment and which workflows need their links and quality fixed first.
If you want to go deeper
Navigara: Connect Your AI Spend Directly to Your Roadmap — You can review the product description, the founder’s measurement approach, and answers about security and unlinked work. producthunt.com
Token Spend Intelligence · Navigara — Shows how Navigara productizes model routing, cost per unit of work, and roadmap linkage. navigara.com
Usage and Cost API - Claude Platform Docs — Lets you review the scope and endpoints for querying Anthropic organization usage and costs. platform.claude.com
Usage | OpenAI API Reference — Official API documentation for querying OpenAI organization usage and cost data by dimensions such as project. developers.openai.com
Choosing measurement frameworks to fit your organizational goals — Explains why AI usage metrics should be interpreted alongside delivery performance, quality, and developer experience. dora.dev
