Merge Gateway vs Portkey

Why

Start with the asymmetry, because it decides most of the page. Portkey was founded in 2023, ships an MIT-licensed gateway core with 12,848 GitHub stars, and runs as hosted SaaS, self-hosted via npx or Docker, or as a data plane inside your own VPC with official Helm charts and an official Terraform provider; Palo Alto Networks closed its acquisition on 29 May 2026 and is making it the core AI Gateway for Prisma AIRS. Merge Gateway launched on 31 March 2026, is proprietary and hosted only, and its Gateway documentation contains no Docker image, Helm chart, binary, install command or changelog page anywhere in its 403 URLs. On breadth and maturity this is not a contest, and pretending otherwise would waste your time.

The structural difference worth deciding on is credentials and money. Portkey is BYOK-only: requests run on your own provider credentials held as virtual keys, and its pricing page lists no token markup and no credit-purchase model. You pay a platform fee instead — a free Developer tier at 10,000 recorded logs a month, $49 a month for Production with 100,000 recorded logs, then $9 per additional 100,000 requests. Merge inverts that. Managed credentials are the default, Merge is the merchant of record, it consolidates provider invoices into one, and it bills LLM cost plus 5% with no platform, seat or request fee. At $10,000 a month of model spend that 5% is $500 against Portkey's $49 plus overage. But Portkey's meter is recorded logs, not tokens, so high-request, small-payload agent loops inflate it in a way Merge's spend-based fee does not. And if you have no provider accounts and do not want to open them, Merge is the only one of the two that will sell you the tokens.

Everywhere else Portkey is simply further along. It has both simple and semantic caching, prompt management, an MCP gateway with OAuth 2.1, configurable rate limits, retries up to five attempts with 1s/2s/4s/8s/16s exponential backoff, and nested per-target request timeouts. Merge has no cache of its own beyond provider passthrough — semantic caching was announced at launch as arriving "in the next few weeks" and five months later appears on no documentation page — no prompt management, no MCP surface, and timeouts, retries and health-check thresholds are all Merge's to set rather than yours. Observability splits the same way: Portkey exports OpenTelemetry to any OTLP backend, naming Datadog, Grafana and LangSmith, and documents plan-specific retention of 3 days on Developer, 30 on Production and 90 on SaaS Enterprise; Merge's tracing is proprietary and header-based, retained about 24 hours, with no OTLP export, no log drain and no published log-retention window. Catalogue counts settle nothing here, because Portkey's own pages say 250+, 1,600+ and 2,300+ models across 40+ to 48 providers, against 274 catalogue rows spanning 24 model-family prefixes counted on Merge on 2026-09-02.

So the honest question is not which is better but whether the narrow case for Merge applies to you. There are two. The first is the Embedded Routing Stack: each of your own customers gets their own routing policies, BYOK credentials, budgets and API keys, with pass-through usage reporting and a strict BYOK_ONLY mode that never falls back to Merge's keys. Portkey has workspaces, virtual keys and per-workspace limits, but per-customer cost splits are not explicitly stated. The second is spend enforcement on a self-serve plan: Merge blocks with HTTP 402 before any provider is called, across org credit, project budgets over five periods, per-key limits with daily, weekly or monthly resets, 50/80/90% alert thresholds and a 10x-daily-average velocity alert, whereas Portkey's docs place granular budget and rate limits on Enterprise in one place and "Enterprise plan and select Pro customers" in another. If neither of those describes you, buy Portkey.