Llm Cost Tracking Bundle Laravel Package

Getting Started

lunetics/llm-cost-tracking-bundle is a Symfony bundle (not a Laravel package) that tallies token usage and computes per-model LLM costs, with a Web Profiler panel for inspecting them. In Laravel you get the cost-tracking core, but not the profiler UI — plan to build your own surfacing (logs, Telescope, a dashboard).

1. composer require lunetics/llm-cost-tracking-bundle
2. There is no package auto-discovery. Register the services you need in a small LlmCostServiceProvider, or instantiate the tracker directly from the container.
3. The entry point is CostTrackerInterface. In v0.6.0 this interface changed (see Gotchas) — read its signature in the source before writing your first implementation.
4. Locate the model pricing map in the bundle config and copy it to config/llm-cost-tracking.php so you can maintain prices yourself.

First use case: a thin decorator around your OpenAI/Anthropic client that records model, prompt_tokens, completion_tokens on each response and returns a computed cost you can log alongside the request.

Implementation Patterns

• Decorator over your LLM client. Don't sprinkle tracker calls through controllers. Wrap your existing client (or Laravel AI SDK binding) so every completion is recorded in one place, including failures.
• Bind the interface, not the class. In your service provider: $this->app->bind(CostTrackerInterface::class, ...). This keeps you insulated from the v0.6.0 interface break and lets you swap in a null/no-op tracker in tests.
• Implement the skipped-call surface. v0.6.0 adds reporting for skipped calls — calls the tracker could not price (unknown model, missing usage in a streamed response, malformed payload). Your implementation must return these; in Symfony they render in the profiler panel, in Laravel you should pipe them into Log::warning() or a metrics counter.
• Per-request and per-job scoping. Cost accumulates across a long-lived worker. Reset/flush the tracker at the start of each queued job and persist the total to your own llm_costs table (user, tenant, job id, model, tokens, cost).
• Config layering. Keep a base price table in config and override via env for premium/discounted contracts; store prices as integers (micro-dollars) to avoid float drift in aggregates.
• Testing. Bind a fake CostTrackerInterface that records calls and returns a fixed cost; assert against that in feature tests rather than mocking HTTP.

Gotchas and Tips

• Breaking change in v0.6.0: CostTrackerInterface changed. Any custom implementation (including test doubles and decorators) will fatal until it implements the new skipped-call methods. Check PR #12 for the exact signature before upgrading; this is a minor-version bump with a BC break, so pin ^0.5 if you can't migrate immediately.
• The profiler panel is Symfony-only. The headline v0.6.0 feature ("surface skipped calls in the profiler panel") gives you nothing in Laravel out of the box — retrieve the skipped-call data from the tracker API yourself.
• Skipped calls are the debugging gold. A spike in skipped calls usually means: a model name missing from the pricing map, a streamed response with no usage block, or an error path recorded without tokens. Alert on it.
• No auto-discovery / no Laravel config merge. You own publishing, config merging and service wiring. Expect to write ~30 lines of glue.
• Double counting. Retries, streaming re-sends, and tool-call loops inflate totals. Key each recorded call by a request/response id and dedupe on write.
• Pricing drift is on you. The bundled table goes stale after provider price changes; treat it as a seed and keep your own overrides under version control.
• Extension point: decorate CostTrackerInterface rather than extending the concrete tracker — it survives upstream class renames and is the cleanest place to add tenant tagging or metrics export.