Where Outlay fits — and where it doesn't.
Teams usually weigh Outlay against four other things: a cloud FinOps suite, an LLM observability tool, the native provider console, or building it themselves. Here's the honest version of each.
Outlay is the last column — scroll to compare →
| FinOps suite | LLM observability | Native console | Outlay | |
|---|---|---|---|---|
| Spend mapped to tickets / epics / roadmap | ✕cloud resources, not eng work | ✕traces & tokens, not scope | ✕per-key, per-model totals | ✓attributed to the work you plan |
| Cache-aware costing (per token class) | ~reads the bill, can't attribute | ✕tokens at a flat rate | ~provider total only | ✓cache reads at ~0.1× |
| Reconciled to the provider invoice | ✓it is the bill | ✕ | ✓single provider | ✓Anthropic, Bedrock, Vertex, OpenAI |
| Forecast by open scope + budget guardrails | ✓for cloud, not AI-by-scope | ✕ | ✕ | ✓quarter forecast + pace alerts |
| Enforce a hard cap by program (block / route-down) | ✕reports the bill, can't act | ✕observes, can't act | ~per-seat caps only | ✓opt-in gateway, by program |
| Prompts never leave your environment | ✓doesn't see prompts at all | ✕ingests prompts/outputs to trace | ✓ | ✓read-only, metadata only |
| Forecast accuracy back-tested on your own work | ✕ | ✕ | ✕ | ✓leave-one-out, with sample size |
| Prompts & keys never leave your environment | ✓reads billing, not content | ✕traces prompts/outputs | ✓ | ✓read-only metadata |
Generalizations across categories — individual products vary. The point isn't that these tools are bad; it's that they answer a different question than "what is each epic costing us in AI, and what do we do about it?"
FinOps suites (Cloudability, CloudHealth, Vantage, Kubecost…)
Excellent at cloud cost — compute, storage, Kubernetes — allocated to teams and cost centers. But AI coding-agent spend lives in SaaS invoices and provider APIs, not your cloud bill, and these suites allocate by cloud resource, not by ticket or epic. They'll tell you the line item is growing; they won't tell you which sprint drove it or which model to use instead. Outlay is the AI-native complement: same FinOps discipline, applied to LLM spend by scope of work. Outlay vs. AI cost dashboards, in depth →
LLM observability (Langfuse, Helicone, Datadog LLM, Arize…)
Great for debugging quality, latency, and traces — and several show cost per call. The structural trade-off: to trace, they ingest your prompts and outputs, so your content sits in their data path. They also stop at observation — they surface cost, they don't forecast by roadmap scope or estimate planned work. And because they cost from raw token counts, a tool that prices cache reads at the base input rate overstates a cache-heavy agentic workload several-fold — on our own usage the gap was 7.4×. Outlay costs each token class separately and reconciles to the provider invoice; it connects read-only (only a category, token counts, and ticket id ever reach us) and is built to act on the number — forecast and budget it — not just chart it.
Native provider consoles (Anthropic, OpenAI, Cursor admin…)
Free, accurate, and already there — totals by API key, model, and workspace. The gap is the join: a key or workspace isn't an epic, and there's no forecast, no pace guardrail, and no cross-tool view (Claude Code + Cursor + direct API in one place). Outlay reads these consoles read-only and does the attribution and forecasting on top.
Build it yourself
Entirely doable — a script that pulls the Anthropic admin API, joins PRs to tickets, and charts it. Most teams who try it find the join is the hard part (detached-HEAD CI agents, tracker teams, branches not named after tickets), and that a one-off dashboard rots. Outlay ships the attribution join with fidelity tiers, the cache-aware cost model, and the forecast as a product — read-only, metadata-only, free during the pilot. If it's not worth more than a weekend to you, build it; if AI is a real and growing line item, don't maintain it yourself.
What about AI gateways & routers? (OpenRouter, Portkey, Martian…)
Different category: gateways are a hosted layer your traffic flows through to reach many models. Outlay isn't in your traffic path at all — it connects read-only to your tracker and your provider's usage data, so your prompts and keys never reach us. Run a gateway if you need one API across many providers; Outlay attributes and forecasts the spend regardless of how your calls are made, and maps it all to the roadmap.
See it on your own numbers.
Two weeks, read-only, prompts never leave your environment. We'll map your real AI spend to your roadmap and show you where it's about to go over.