# Portkey Alternatives: Gateways, Routers, and Flat-Rate Lanes

Portkey bundles a gateway, observability, and guardrails. The real alternatives sorted by job: LiteLLM, OpenRouter, Langfuse, Helicone, and a flat-rate lane for OpenAI spend.

*Published 2026-06-12 · https://proxyllm.ai/blog/portkey-alternatives*

Portkey is three products in one gateway: routing across your provider keys, observability with logs and traces, and guardrails that police what goes in and out. The right alternative depends on which of those jobs you are actually replacing. LiteLLM for the self-hosted router, OpenRouter for hosted model breadth, Langfuse for observability on its own, Helicone with a caveat, and, when the job is the bill itself rather than the plumbing around it, a flat-rate lane like ours. ProxyLLM is our product; the disclosure applies wherever it appears below.

## What Portkey is, fairly

Portkey deserves an accurate description before the alternatives line up. It is a hosted AI gateway with an open-source core: your requests route through it using your own provider keys, and the platform adds request logs, traces, cost attribution, guardrails, semantic caching, and fallback routing, priced as SaaS plans while inference bills at provider cost. For teams under compliance pressure that need policy enforcement in the request path, it is one of the few tools that bundles the whole control plane.

If that bundle is exactly what you need, keep it. Alternatives make sense when you are paying for three jobs and using one.

## If routing and fallbacks are the job

Two directions, depending on who should operate the router.

**LiteLLM**, self-hosted: an open-source proxy speaking the OpenAI format across 100+ providers, with virtual keys, per-key budgets, rate limits, and fallback chains you configure. No gateway fee, your infrastructure, your pager. It is the standard answer when the requirement says the request path stays in-house.

**OpenRouter**, hosted: one key against 400+ models with marketplace failover and a fee of roughly 5.5% on credits as of June 2026. Zero operations, maximum breadth. The head-to-head between these two philosophies is in [LiteLLM vs OpenRouter](/blog/litellm-vs-openrouter), and the wider field is in [OpenRouter alternatives](/blog/openrouter-alternatives).

## If observability is the job

**Langfuse** is the strongest standalone pick: open source, self-host or cloud, with traces for multi-step agent runs, evals, and prompt management. It instruments through SDKs rather than sitting in the request path, which suits teams that want observability without adding a network hop.

**Helicone** historically competed here with proxy-style one-line setup; after its 2026 acquisition by Mintlify the product kept running, with roadmap questions covered in [Helicone alternatives](/blog/helicone-alternatives).

Lightweight logging also comes built into gateways now, ProxyLLM included: our endpoint logs every request with the model, the serving lane, and its API-equivalent dollar value, on a $0 tier. A built-in log is not a tracing platform, and for plenty of teams it is the entire requirement.

## If guardrails are the job

Honesty favors the incumbent here. Gateway-level guardrails, input and output checks enforced in the request path across every app, are Portkey's strongest and least replaceable feature. The alternatives are architectural rather than direct: validation libraries and moderation checks in application code, which work well per-app and poorly as org-wide policy, or accepting the gateway as your enforcement point and keeping Portkey for exactly that.

If guardrails are the core requirement, Portkey is hard to replace, and an alternatives page that pretends otherwise is selling you something.

## If the bill is the job

Every tool above organizes spend. None of them reprices it, because all of them sit on per-token meters. Swapping gateways rearranges a bill; changing the cost model is what shrinks it.

That is the lane we built, disclosure repeated: ProxyLLM's [Codex Hosted](/) runs OpenAI's official, unmodified Codex CLI signed in with your own ChatGPT account (you run the login command on your own machine), served as an OpenAI-compatible endpoint. OpenAI-bound volume bills to the flat subscription, $129 a month with no inference markup on top. Predictability is the actual product: the OpenAI line item becomes a fixed number, and by our planning estimates a Pro 5x plan absorbs roughly $3,500 of API-equivalent work a month, so a $3,500 metered month maps to a $100 plan plus that flat $129. Estimates, never guarantees. The caveats, stated plainly: the flat lane serves OpenAI models only, and it returns complete responses rather than streams, so streaming UIs belong on a metered key lane.

## The table

| Job you hired Portkey for | Hosted pick          | Self-hosted pick    | The trade                                   |
| ------------------------- | -------------------- | ------------------- | ------------------------------------------- |
| Routing and fallbacks     | OpenRouter           | LiteLLM             | Marketplace fee vs your own ops             |
| Observability and traces  | Langfuse Cloud       | Langfuse, self-host | SDK instrumentation vs proxy convenience    |
| Lightweight request logs  | ProxyLLM ($0 tier)   | LiteLLM callbacks   | A log, deliberately, not a tracing platform |
| Guardrails in the path    | Portkey (keep it)    | App-level checks    | Org-wide policy vs per-app code             |
| A flat OpenAI bill        | ProxyLLM ($129 flat) | DIY Codex proxy     | OpenAI-only lane, complete responses        |

Every row has a winner that is not us, except the ones where the job is the one we built for. That is how an honest alternatives table should read.

## Choosing in one pass

Name the job, then pick the column you can operate. Teams with platform engineers and compliance requirements lean self-hosted; teams that ship product lean hosted. And if the reason you opened this page is that the OpenAI line keeps growing no matter which gateway fronts it, that is a cost-model problem, mapped across every workload in [the best LLM gateways in 2026](/blog/best-llm-gateway-2026).

For the cost-model version of the question, the [calculator](/calculator) takes your current monthly OpenAI spend and shows the flat-lane arithmetic in about thirty seconds.