Cost vs Portkey: an honest comparison

Portkey is a serious production gateway, and it is stronger than Cost everywhere that role matters. One API reaches roughly 1,600 models across 45-plus providers, with conditional routing, automatic fallbacks, key-level load balancing, retries with a circuit breaker, canary rollouts, and both simple and semantic caching. If you want a resilient control plane in front of your models, Cost does not compete with that, and is not trying to.

It is also credibly open source. The gateway is MIT-licensed with around 12k GitHub stars, and Portkey moved formerly paid governance, observability, and cost-control features into the open repo. You can self-host the full stack for free, and enterprises can run it in a VPC or fully on-prem, which addresses the data-residency objection that a pure-SaaS tool cannot.

The governance surface is broad: RBAC, SSO, activity logging, PII redaction, and a large library of deterministic and LLM-based guardrails, plus a recently shipped MCP Gateway for governing agent tool access. For a platform team standardizing how an organization talks to models, that breadth is genuinely useful.

And the acquisition by Palo Alto Networks in 2026 is a real signal. It validates enterprise trust and adds durability and security weight behind the product. If you are buying for a large enterprise that already works with Palo Alto, that is a point in Portkey's favor, not against it.

Cost verifies before it recommends, which Portkey does not do. Portkey gives you routing, fallbacks, and budgets, but if you want to move a route to a cheaper model, the decision is still yours to make blind. Cost replays your recent real traffic through the cheaper model, judges every response across five dimensions, and only marks the swap safe if 95% pass. You get the saving with proof attached.

Cost is out-of-band, and for many teams that is the whole point. Portkey is an inline gateway, so every request passes through it and it sits in your latency-critical path. Cost is an SDK that wraps your client where it already runs. There is no extra hop, no new point of failure between your app and the provider, and nothing to keep highly available.

Cost is privacy-first by default without operational overhead. It sends only cost metadata, model, tokens, route, and latency, and keeps prompt and response bodies in your network unless you opt in per route. Portkey can meet a strict data-residency bar too, but you usually get there by self-hosting the gateway or running a VPC deployment, which is more to operate.

And Cost is cost-shaped. The unit is the euro, attributed to a route or feature, and the deliverable is a ranked list of what to fix with expected savings. Portkey's cost features are budgets, rate limits, and per-use-case attribution inside a gateway. Cost is built for the person who has to justify and reduce the bill, and the verification step is the part they can take to a meeting.

If you are a platform or infrastructure engineer who needs one resilient entry point to many models, with routing, fallbacks, caching, and guardrails, Portkey is the right tool and Cost is not an alternative to it. You are buying a gateway, and Cost does not provide one. Choose Portkey and run it inline.

If you are a CTO, a head of engineering, or a CFO whose problem is a line item that keeps climbing, Cost is built for your question. You do not want to insert a gateway into the request path just to understand spend. You want to know which routes cost the most, whether they can run cheaper safely, and proof before anyone changes a model. Cost attributes the spend and runs the verification without touching your hot path.

Latency and risk tolerance often decide it. An inline gateway is a powerful pattern, but it is also a dependency on every request, and some teams will not accept a third party between their app and the provider. Cost's out-of-band design sidesteps that entirely, which makes it an easier security and reliability conversation.

The acquisition is worth weighing honestly on both sides. For an enterprise already standardized on Palo Alto, Portkey under Prisma AIRS may be a strategic fit. For a team that wants a focused, independent cost tool and is wary of a product being folded into a larger security suite, Cost's independence and narrow scope may be the safer bet. Either way, the two are not mutually exclusive.

Yes, and they slot together cleanly. Portkey can be your gateway on the request path, handling routing, fallbacks, caching, and guardrails. Cost can sit out-of-band as the SDK that attributes spend per route and runs verified downgrade recommendations. Neither needs the other to be removed.

Running both lowers the cost of adopting Cost. Your platform team keeps the gateway they standardized on, and your leadership gets per-route attribution and a safety gate on model changes. Portkey decides how requests are routed; Cost decides which model each route should be on, and proves it before you switch.