An AI gateway routes models. You do not need one to see OpenAI cost per customer.

An AI gateway sits in front of many models. The app talks to the gateway. The gateway picks a provider, may cache the answer, may issue a virtual key, and often keeps a log of the prompt. The pitch is control: one endpoint, many backends, budgets, fallbacks.

You do not need that box to see OpenAI cost per customer. Cost per customer is a tag on the request and a report that puts cost next to plan price. Routing does not create that report. Cache does not create it. A virtual key does not create it. You do not need an AI gateway for the P&L.

CostSlice is not an AI gateway. It is a pass-through OpenAI proxy plus a customer P&L. OpenAI only. You keep your OpenAI API key. We forward Authorization. We do not replace it. We do not route. We do not cache. We do not fall back to another model. We do not store prompt bodies. We count, then show which customer is unprofitable.

Fair one-liners, because the category is crowded. Helicone logs and can route — see vs Helicone. Portkey issues virtual keys through a catalog — see vs Portkey. LiteLLM unifies providers and can budget-stop calls — see vs LiteLLM. OpenRouter is a many-model endpoint you pay through them. Kong is an API gateway that can grow AI plugins. Cloudflare can sit at the edge with AI products of its own. None of those sentences are a ranking. They are what the box is for.

Why still want a gateway? You route across providers. You want a semantic cache. You want a virtual key so the app never holds the provider secret. You want fallbacks when a model is down. You want a log of what people typed. Those are gateway jobs. Buy a gateway for them. Do not buy a gateway only to learn that acme-health is −101%.

Gateways often store prompts. That is how they debug and how they cache. It is a choice. CostSlice’s choice is the opposite. Bodies never hit our disk. Usage metadata is enough to price a tenant. If your review board will not let a middle box keep text, start with the meter, not the gateway.

How the meter works: point base_url at https://api.costslice.com/v1. Send x-cs-key. Tag x-cs-tenant, x-cs-feature, x-cs-env. Completions still go to OpenAI. Missing tags do not drop the call. We do not cap a tenant because a row is red.

Starter is a hard cap on the CostSlice plan, not a gateway budget. 10,000 requests. No overage. No card. Pro and Scale meter overage at $6 per 100k. They do not shut a customer off.

Claude is not on CostSlice today. A gateway that sells many models will mention Claude on day one. We do not. OpenAI only. Other models come later.

Starter $0 / Pro $79 / Scale $199, overage $6/100k. No credit card on Starter. Pricing on the homepage wins if a number here is stale.