Portkey

Portkey sits between your application and the model providers it calls. You point an OpenAI-style client at the Portkey gateway, and every request to OpenAI, Anthropic, Google, Azure, Bedrock or one of 1,600+ other models goes through one place where it can be routed, retried, cached, logged and checked.

What the gateway does. Fallbacks, load balancing and automatic retries keep a request alive when one provider degrades. Simple caching is on every plan, and the paid plan adds semantic caching, which answers near-identical prompts from cache instead of paying for them again. Guardrails run deterministic checks on the free tier and LLM-based or partner guardrails on paid plans, and the gateway also fronts MCP servers and agent traffic, not just chat completions.

Seeing where the money goes. Logs, traces, custom metadata and filters show cost, latency and usage per request, per team or per use case, and an enterprise tier adds granular budgets and rate limits per key. A model catalogue lists input, output and cache prices for each provider's models side by side.

Prompts. Templates are versioned, take variables, can be tested in a playground and are served from API endpoints, so prompt text does not have to be hard-coded in the app.

Open source and ownership. The core gateway is published as the MIT-licensed @portkey-ai/gateway package and can be self-hosted. Palo Alto Networks announced in 2026 that it is acquiring Portkey, and the hosted product now also appears as the Prisma AIRS AI Gateway.

Pricing. The free Developer plan records 10,000 logs a month with 3-day log retention and is meant for prototyping. Production is $49 a month for 100,000 recorded logs (then $9 per extra 100,000), 30-day retention, alerts, role-based access and semantic caching. Enterprise is custom-priced, with private cloud or VPC deployment, SSO, data residency and SOC 2, GDPR and HIPAA compliance.