Portkey Review (2026)

What Portkey actually is

Portkey is an LLM gateway. It sits between your app and your model providers as a drop-in proxy, routing calls to 1,600+ models through one API. On top of routing it adds fallbacks, load balancing, retries, timeouts, caching, budgets, rate limits and 50+ guardrails - plus an observability backend with logs, traces and metrics, and prompt management.

That makes it a different animal from the other tools in this category. The others are eval-first or trace-first. Portkey is a control plane for spend, routing and governance that happens to include observability. If your problem is “we call five providers, costs are exploding, and we need fallbacks and budgets in one place,” Portkey is aimed squarely at you. If your problem is “we need to score whether our agent’s answers are good,” that’s a different tool.

It’s a well-backed one. Portkey was founded in January 2023 by Rohit Agarwal and Ayush Garg, and raised a $3M seed followed by a $15M Series A led by Elevation Capital with Lightspeed. The open-source gateway repo sits at roughly 12,500 GitHub stars.

The distinctive part: the gateway itself

The gateway is the product, and it’s fast and broad. One base URL change routes your existing OpenAI SDK calls through Portkey, auto-enriched with provider config, caching status, retry attempts and prompt versions. No re-instrumentation for the routing path.

Portkey markets it as sub-1ms added latency with a 122kb footprint, handling 10B+ requests a month for 24,000+ orgs. Those are vendor numbers. A benchmark published by Kong, a competing gateway vendor, reported Portkey at about 65% higher latency than Kong’s own gateway under equivalent conditions. Both parties are interested, so both figures are claims - the honest read is that Portkey’s latency is competitive but not independently settled, and you should test on your own path.

The gotcha: observability is gated

This is the thing to understand before you plan a deployment. The open-source gateway gives you routing and a basic dashboard. Real observability lives on the paid tier.

The Apache 2.0 gateway self-hosts free and does the hard networking - routing, retries, fallbacks, load balancing, guardrails. What it does not give you is meaningful logging, traces, analytics or retention. For that you move to managed Production, or you build your own logging stack around the gateway. So a team that self-hosts Portkey expecting a full observability platform gets a proxy and a basic dashboard, and discovers the analytics they wanted are a paid feature.

That’s not dishonest - it’s open-core, and the free part is genuinely useful. But it inverts the usual self-host promise in this category. With Langfuse, self-host gets you the whole product. With Portkey, self-host gets you the pipe, not the dashboard.

Pricing: watch the log meter

Tier	Price	Logs/mo	Log retention	Overage
Open Source	$0	self-hosted	-	Apache 2.0
Developer	$0	10,000	3 days	none
Production	$49/mo	100,000	30 days	$9/100k requests
Enterprise	Custom	10M+	configurable	-

One detail bites. The meter caps logs, not requests. Past the cap, your requests keep routing normally - you just stop recording logs beyond the limit unless you pay the overage, which is $9 per additional 100k requests. That’s friendlier than a hard request cap in one sense (your app never breaks), but it means your observability quietly goes dark at volume unless you’re watching the meter. There’s no annual discount published.

Portkey versus a pure observability tool

If you’re comparing Portkey to Langfuse or Laminar, you’re comparing different jobs. Portkey’s core value is routing, fallbacks, budgets and guardrails across many providers - the gateway. Its observability is a bundled bonus that gets good on the paid tier. A pure observability tool’s core value is the tracing and eval depth, and it doesn’t route your traffic. Plenty of teams run both: Portkey as the gateway, and a dedicated observability backend fed over OpenTelemetry. Portkey’s strong OTel support, enriching traces with cost and token metrics per the GenAI conventions, makes that pairing straightforward.

Should you use it?

Use Portkey if you call multiple LLM providers and want one place to route, cache, budget, rate-limit and guardrail them, and you’re fine paying for the managed tier once you need real logs and analytics.

Don’t use Portkey if you want free, fully self-hosted observability - the OSS build won’t give you that - or your only need is eval depth, where a purpose-built eval tool serves you better.

Bottom line: the best gateway in this set, with a genuinely useful free open-source core. Just map the split before you deploy - routing is free and self-hostable, but the observability most teams want is on the paid tier.

---

Pricing and features verified against portkey.ai on 23 July 2026. This category ships breaking changes monthly - we re-verify every 30 days.