Short answer. LiteLLM is an MIT-licensed gateway you run: virtual keys, budgets, routing, guardrail hooks, logging and an OpenAI-compatible endpoint. Portkey is now Prisma AIRS AI Gateway under Palo Alto Networks: an enterprise control plane with an open-source gateway core, a much larger connector catalogue and a sales-led contract. Choose LiteLLM for self-hosting and cost control; choose Portkey/Prisma AIRS for a vendor-supported security and governance plane.
The architectural difference
LiteLLM started as a Python SDK that normalizes provider APIs and grew into a proxy. The proxy reads a config.yaml, exposes an OpenAI-compatible endpoint, and adds virtual keys, budgets, rate limits, routing, caching, guardrail hooks and logging callbacks. It is designed to run in your VPC, including air-gapped environments on the Enterprise plan.
Portkey started as an AI gateway and expanded into observability, guardrails, prompt management and an MCP gateway. Its site now says Portkey is Prisma AIRS AI Gateway, generally available as part of Palo Alto Networks. The open-source gateway remains on GitHub under the MIT license; the commercial product is sold through Palo Alto Networks, and its pricing page redirects to a contact-sales form.
That is the real fork in the road. LiteLLM is a product you install and operate. Portkey is a control plane you buy, with an open-source data plane you can run. The feature overlap is large; the operating model is not.
Feature comparison
| Capability | LiteLLM | Portkey / Prisma AIRS |
|---|---|---|
| License | MIT for the gateway and SDK | MIT for the open-source gateway; commercial control plane |
| Deployment | Self-hosted, VPC, air-gapped on Enterprise | SaaS, hybrid or self-hosted depending on contract |
| Provider coverage | 100+ providers via the SDK and proxy | 1,600+ LLMs via a universal API |
| Routing | Fallbacks, load balancing, retries, budgets, RPM/TPM limits | Fallbacks, conditional routing, load balancing, canary, circuit breaker |
| Guardrails | Provider integrations; pre-call, post-call, during-call and logging-only modes | 50+ guardrails, PII redaction, policy enforcement |
| Observability | Logging callbacks, spend logs, request metadata | Built-in dashboards, traces, cost and usage analytics |
| Identity | Virtual keys; SSO, SCIM and audit logs on Enterprise | SSO, RBAC, audit logs on enterprise plans |
| Pricing | Open source free; Enterprise quote | Quote; no public self-serve price after the acquisition |
Feature labels are a planning summary. Portkey's product naming and packaging changed with the Palo Alto Networks acquisition; verify the current SKU and deployment terms before purchase.
Cost
LiteLLM's open-source gateway has no license fee. The bill is compute, database, observability storage, egress and the engineer who owns upgrades and incidents. A small gateway can run on a modest instance; a multi-tenant deployment with high availability, Redis, Postgres and log storage is a real service. LiteLLM Enterprise is quote-based, with a 30-day trial key and a 24/7 support SLA.
Portkey/Prisma AIRS is also quote-based. After the acquisition, there is no public self-serve price to compare. That does not make it expensive or cheap; it makes it unpriceable without a sales conversation. Ask for the metric (requests, tokens, seats, environments), the minimum term, the overage and the support tier.
gateway_tco =
replicas × instance_hourly × 730
+ database and Redis
+ observability storage and retention
+ egress
+ license or enterprise subscription
+ engineering time
# The license line is only one line.
# A cheap gateway with a fragile upgrade path is not cheap.
Latency overhead
Every gateway adds a hop. The proxy itself is usually a small fraction of total latency; guardrails, logging and retries are the larger costs. Palo Alto Networks claims less than 1 ms p50 overhead for Prisma AIRS AI Gateway. Treat that as a vendor claim and measure your own p95 with streaming and the guardrails you will actually enable.
LiteLLM publishes benchmarks and exposes timing in its logs, but the same rule applies. A 200 ms PII scan dominates a 1 ms proxy. If you run guardrails in logging_only mode, LiteLLM can keep the scan off the response path; if you run a blocking pre-call check, the latency is part of the product. Test the configuration you will ship, not the default.
Guardrails and governance
LiteLLM's guardrail system supports pre-call, post-call, during-call and logging-only modes, with integrations for providers such as Presidio, Bedrock, Lakera, Aporia and generic guardrail APIs. Policies can be scoped to teams, keys or models on Enterprise. The open-source proxy already exposes the hook, which is useful for a pilot; the enterprise controls are what make it auditable across an organization.
Portkey's guardrails are a first-class product: 50+ integrations, PII redaction, policy enforcement and observability in one control plane. The trade-off is that you are buying a security vendor's plane, not just a proxy. That can be the right call for a regulated enterprise that wants one throat to choke for AI traffic, and the wrong call for a small team that only needs a base URL and a budget.
For either product, verify the security claims in the contract. SOC 2 Type 2 and ISO 27001 are vendor statements until you see the report scope and date. Ask about data retention, prompt logging, subprocessors and support access. This page is not legal or security advice.
Migration
Both are OpenAI-compatible, so the client change is a base URL and a key. The gateway configuration is the work.
# LiteLLM config.yaml (illustrative)
model_list:
- model_name: gpt-oss-120b
litellm_params:
model: openai/gpt-oss-120b
api_base: https://api.plugsky.com/v1
api_key: os.environ/PLUGSKY_API_KEY
general_settings:
master_key: os.environ/LITELLM_MASTER_KEY
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:4000",
api_key="LITELLM_KEY",
)
# Point the same client at Portkey/Prisma AIRS by
# changing the base URL and key, not the application.
- Inventory routes. Model aliases, fallbacks, retries and timeouts are the migration surface.
- Map keys and budgets. Virtual keys, teams and spend limits do not translate one-to-one.
- Re-create guardrails. Provider, mode and policy scope must be tested, not assumed.
- Re-point logging. Trace IDs, retention and PII handling are part of the contract.
- Run both gateways in parallel on a sample and compare latency and error rate.
- Cut over by route. Keep the old gateway warm for rollback.
If you are choosing a gateway for the first time, read what an LLM gateway does before comparing vendors. For the alternatives, see LiteLLM alternatives and Portkey alternatives.
When to choose
- LiteLLM: you want an MIT-licensed gateway you can run in your own VPC, with budgets and routing under your control.
- LiteLLM Enterprise: you need SSO, SCIM, audit logs, air-gapped deployment and a 24/7 SLA without buying a security platform.
- Portkey/Prisma AIRS: you want a vendor-supported control plane with a large connector catalogue, guardrails, observability and enterprise support.
- Neither: you use one provider and one key and have no policy requirement. A gateway is unnecessary complexity until you need routing or central controls.
If the gateway is only there to reach models, the model layer can be simpler. Plugsky provides an OpenAI-compatible API with 39 models and flat plans; see pricing. It is not a replacement for LiteLLM's budget or guardrail controls.
Sources
- LiteLLM documentation — checked 20 November 2026.
- Portkey documentation — checked 20 November 2026.
- LiteLLM repository and license — checked 20 November 2026.
FAQ
Is LiteLLM open source?
Yes. LiteLLM gateway and SDK are MIT-licensed. The OSS proxy includes virtual keys, budgets, rate limits, routing, caching, guardrail hooks and logging. SSO, SCIM, audit logs, air-gapped support and 24/7 SLAs are Enterprise.
Is Portkey still available as a standalone product?
Portkey's site now says it is Prisma AIRS AI Gateway, generally available under Palo Alto Networks. Its pricing page redirects to a contact-sales form. The OSS Portkey gateway remains MIT on GitHub. Confirm which commercial product applies.
Which is cheaper, LiteLLM or Portkey?
LiteLLM OSS has no license fee; you pay infrastructure and engineering. LiteLLM Enterprise and Portkey/Prisma AIRS are quote-based, and Portkey no longer publishes a self-serve price. Compare gateway compute, storage, egress, support and engineers.
Which adds less latency?
Both add a hop, and guardrails add more. Palo Alto Networks claims under 1 ms p50 overhead for Prisma AIRS. Treat that as a vendor claim and measure your p95 with streaming and the guardrails you will enable.
Can I self-host both?
LiteLLM runs in your infrastructure, including air-gapped on Enterprise. Portkey has an OSS gateway you can self-host; its commercial control plane is SaaS, hybrid or self-hosted by contract. Verify feature parity.
Do I need an LLM gateway at all?
If you use one provider, one key and no guardrails, a gateway is overhead. If you route across providers, need per-team budgets, logging or policy enforcement, it is cheaper than rebuilding those controls.
Reviewed 20 November 2026. LiteLLM's licensing, enterprise features and support terms were checked against its public pages; Portkey's product naming and pricing path were checked against portkey.ai on that date. Acquisition terms, security reports and contracts require direct verification. This page is editorial and does not rank providers by payment.