Appearance
Foundation APIs
This is the layer most often missed in AI spend tracking. Foundation API spend lands in a billing account — AWS, Azure, an OpenAI org account — rather than as a SaaS line item, so finance never sees it as "AI spend" and engineering never reports it because they think of it as "infra." All usage-based, all invisible by default unless someone goes looking.
What's tracked
| Tool | Vendor | Pricing model | Coverage today | Notes |
|---|---|---|---|---|
| OpenAI API | OpenAI | Per-token | Catalog-listed · Hostname-recognised · Billing-connectable (via OpenAI Codex connector) | The Codex connector polls the same OpenAI billing surface — covers API usage too. |
| Anthropic API | Anthropic | Per-token | Catalog-listed · Hostname-recognised · Billing-connectable (via Claude Code connector) | Same — Claude Code admin key reads org-wide billing. |
| Google Gemini API | Per-token via Vertex / AI Studio | Catalog-listed · Hostname-recognised · Billing-connectable | Gemini connector. | |
| AWS Bedrock | Amazon | Per-token, billed via AWS | Catalog-listed | Lands on the AWS bill — no Flowstate connector to AWS billing. Manual entry. |
| Azure OpenAI Service | Microsoft | Per-token, billed via Azure | Invisible | Lands on the Azure bill. Manual entry. |
| Together AI | Together AI | Per-token gateway | Invisible | Manual entry. |
| Fireworks | Fireworks | Per-token gateway | Invisible | Manual entry. |
| Groq | Groq | Per-token (fast inference) | Invisible | Manual entry. |
| Replicate | Replicate | Per-token / per-second gateway | Invisible | Where Stable Diffusion / Flux usage often lands. |
| fal.ai | fal.ai | Per-second gateway | Invisible | Same — image / video model serving. |
| OpenRouter | OpenRouter | Per-token gateway across providers | Invisible | Manual entry. Multi-model routing makes attribution hard. |
What Flowstate misses today
The two biggest gaps: AWS Bedrock and Azure OpenAI Service. Both are usage-based, both bill through cloud accounts, and neither has a Flowstate connector. If your engineering org is using either at scale, the spend is hiding inside your AWS or Azure invoice — go pull the line items and enter them as AI Agents per project.
The gateway providers (Together, Fireworks, Groq, Replicate, fal.ai, OpenRouter) are the third gap. Engineering teams reach for these to get cheap or fast inference; the spend looks small per call but compounds quickly. Model each used gateway as its own AI Agent with the monthly invoice value.
Conversely: if you've already wired the OpenAI Codex, Claude Code, or Gemini billing connectors, your foundation API spend on those vendors is already covered. The connector reads org-wide billing — chat seats, code agents, raw API calls all roll into the same number.