Per-Request LLM Cost Attribution: A FinOps Playbook

Granular LLM cost attribution moves AI spend from a monthly bill to an operational metric. This enables product decisions based on unit economics and customer profitability.

The FinOps Foundation's 2025 State of FinOps report indicates 63% of respondents now manage AI spending, a near doubling from 31% the prior year. This shift signals AI cost becoming a primary FinOps concern, moving beyond a side bucket in cloud spend. For companies spending $5,000 to $50,000 monthly on LLM APIs, aggregate vendor bills fail to explain cost spikes or feature-specific profitability.

Opaque LLM Spend Hinders Product Decisions

Monthly vendor invoices from providers like OpenAI or Anthropic offer a total spend figure but lack the granularity to answer critical questions. Engineering teams cannot identify which feature drove a cost spike, nor can finance attribute spend to specific teams or customers. This opacity prevents understanding the unit economics of AI-powered features, making it difficult to assess profitability or the impact of prompt changes. The post argues that per-request attribution transforms AI spend from a monthly surprise into an actionable operational metric.

Minimum Schema for Cost Tracking

Implementing per-request attribution does not require a complex data platform, but it demands a disciplined event schema. The post outlines a minimum set of fields for each LLM request record: timestamp, provider, model, input_tokens, cached_input_tokens, output_tokens, request_id or trace ID, team, feature, customer_id or workspace ID, environment (e.g., prod, staging), and status (e.g., success, timeout, retry, fallback). Without feature, team, and customer_id, the ability to perform margin analysis and allocate ownership is lost. The status field is critical for distinguishing genuine demand from silent cost inflation due to retries.

Calculating Per-Request Costs

The core formula for calculating individual request cost is straightforward. The author specifies: request_cost = (input_tokens / 1_000_000 * input_rate) + (cached_input_tokens / 1_000_000 * cached_input_rate) + (output_tokens / 1_000_000 * output_rate) + any tool or search fees. The post provides specific pricing examples as of June 8, 2026: OpenAI's GPT-5.4 mini is claimed at $0.75 per 1M input tokens and $4.50 per 1M output tokens. Anthropic's Claude Sonnet 4 is reported at $3 per 1M input tokens and $15 per 1M output tokens. This calculation, applied to every request and enriched with ownership dimensions, allows for rollup into team, feature, and customer-level views.

What We'd Change

Implementing this granular attribution playbook presents practical challenges beyond the schema definition. Integrating team, feature, and customer_id into every LLM API call requires consistent instrumentation across diverse codebases and potentially legacy systems. Many existing applications may not pass these business context tags directly to their LLM wrappers, necessitating significant refactoring or middleware to enrich the request logs. The post acknowledges that gateway logs alone are insufficient without this enrichment.

The dynamic nature of LLM pricing and model versions also complicates long-term cost attribution. While the formula accounts for input_rate and output_rate, these values are not static. Regular updates to pricing tables and the introduction of new models with different cost structures require continuous maintenance of the rate data used in calculations. For smaller teams or those with lower LLM spend (below the reported $5,000 threshold), the engineering overhead of building and maintaining such a system might outweigh the immediate financial benefits. A simpler, aggregate approach with periodic deep dives might suffice until spend scales.

Per-request LLM cost attribution moves AI spend from a line item on a monthly bill to a core operational metric. This visibility enables product teams to make informed decisions about model choice, prompt engineering, and feature design based on actual unit economics. For finance, it provides the data necessary for accurate departmental chargebacks and customer-level profitability analysis. This tactical shift is essential for any product where LLM usage directly impacts margins and product strategy.

The investor read

The increasing focus on per-request LLM cost attribution signals a maturation in the AI product market. Early-stage AI companies often prioritize rapid feature development over granular cost control, but as spend scales and competition intensifies, unit economics become paramount. Products demonstrating this level of operational discipline—tracking costs by feature, team, and customer—present a stronger investment case. This capability indicates a clear path to profitability and defensible margins, moving beyond a simple "AI wrapper" narrative. Investors are looking for teams that can articulate not just what their AI features do, but what they cost and how those costs scale.