Who owns the GPU fleet decides whether LLM routing is a library or a gateway

A drop-in replacement for the OpenAI SDKs means the call shape in your application stays where it is while the library translates to whatever backend answers ^[2]. The dev.to comparison is plain about what comes with that: you manage updates, secrets and scaling, and it is "not a set-and-forget service" ^[3]. LiteLLM is free to download, and the post counts the real price as server costs and developer hours ^[20].

On the rented side, the gateway sits between the application and the vendor APIs. Portkey and Cloudflare AI Gateway are the post's examples, picked for observability and edge performance ^[6]. Vercel AI Gateway routes requests through serverless infrastructure and is aimed at frontend-heavy applications ^[7]. Both managed options are tied to their own ecosystems, and leaving Cloudflare means rebuilding your gateway logic ^[8].

OpenRouter gives one API key for many models, so you can test a new model without opening a separate vendor account ^[10]. It charges a margin on usage. That margin puts the total above the cost of calling the provider directly ^[12]. It also does not expose underlying provider names in its public API the way dedicated gateways do ^[11], so if you have to know which provider served a given request, it will not tell you ^[11].

The $113M Series B at a $1.3B valuation arrives secondhand. The comparison attributes both figures to media reports and leaves out the investors and the date of the round ^[28]. Capital landing on the managed side leaves the deployment question where it was. If the weights run on GPUs you operate, a hosted key does not reach them, and the post's answer for that case is a bridge that abstracts backend differences ^[5].

The free local-first client the post recommends lists 101 models and 36 routing aliases, both counts taken from the tool's own public API endpoint ^[14][15]. The aliases let you ask for "fastest" or "cheapest" instead of naming a model ^[15]. That is about one alias for every three models listed ^[25]. Four preset agent clients ship with it, handling tool calling and memory management ^[16]. It has no observability dashboards of the kind Portkey and Cloudflare provide ^[19], and no public API for custom integrations, so routing logic beyond the presets is not available to you ^[18]. A scorecard that opens by defining provider support as reach and then recommends a tool that does not list its providers has three of its five dimensions left to check ^[26].

What transfers here is the structural match, because all five of the scored dimensions describe the tool and none of them measures your traffic ^[27]. For a corporate data pipeline mixing internal and external models, the recommendation is LiteLLM ^[21]. For a consumer-facing chat app it points to Vercel AI Gateway to keep requests close to the user ^[23], and for audit trails without servers to run, Portkey ^[22].