Ramp opens its internal AI Router to early users

Ramp also says the service applies caching, compaction, semantic attribution and more than 100 other optimisations. That is important context for the 30% figure: the page does not separate the contribution of model routing from those other techniques. It would be misleading to treat the reported reduction as the effect of picking a different model alone.

Routing is an established option, not the whole distinction

The single-endpoint premise predates this launch. In January 2025, OpenRouter said its upgraded Auto Router selected among 19 models and was testing routing powered by Not Diamond’s meta model. That announcement was a product statement at the time, not evidence of its current catalogue or performance; it nevertheless shows that automated cross-model selection was already available.

Ramp’s differentiating pitch is the layer around that selection. It says Token Spend Management can attribute requests by model, product, team and project, and that Router supports current OpenAI, Anthropic and Gemini models plus selected open-source models including Kimi. These are company-described capabilities, not a published comparison of route quality against direct-provider use or competing routers.

That distinction matters commercially. Going direct lets a customer choose the provider, model catalogue and pricing structure itself; Ramp says Router makes those choices dynamically within its eligible set. The benefit is less maintenance when models or prices move. The trade-off is that the customer needs enough visibility into the routing criteria, final model and fallback behaviour to judge whether the delegated decision fits its application.

The commercial terms are still early-access terms

Ramp says Router is free to use at launch and customers will pay list price for the tokens they use. Its first 500 invited users will receive $100 in promotional credits. The page also says no Ramp card, account or LLC is required, but access is initially by invitation and other applicants can join a waitlist.

That is narrower than the “publicly available” wording in a contemporaneous account. Ramp’s own invitation language is the clearer evidence of the present rollout terms.

“Free to use” also is not the same as a demonstrated all-in price advantage. List-price token billing and introductory credits do not reveal how a customer’s routed token bill will compare with direct provider arrangements once credits end. Nor do the published figures show whether savings persist for workloads that cannot accept a lower-cost substitute.

What would settle the customer case

The next decision for an interested team is not simply whether to replace an endpoint. It is whether Ramp can show that its router meets a defined quality threshold on that team’s tasks while preserving useful control over provider and model choice.

Evidence that would make the claim testable includes a disclosed baseline for the 30% internal reduction; workload-level quality, latency and cost results; the share of requests redirected during failures; and a clear account of how customers can constrain or audit the chosen route. Until then, the launch establishes a substantial internal deployment and an early-access product, but not a general proof that delegated routing saves money without changing the output that matters.