Portkey vs LiteLLM compared after the $700M Palo Alto acquisition. Performance, cost, compliance, and what changes now that Portkey is a security product.
Palo Alto Networks announced its intent to acquire Portkey on April 30, 2026, and the deal has since closed. Financial terms were never disclosed; The New Stack characterized it as a "$700M-class AI bet". Portkey will become the AI Gateway for Prisma AIRS. LiteLLM remains independent and open-source.
That single fact reshapes the portkey vs litellm decision for every team running either tool in production. It is no longer a feature comparison between two AI gateways. It is a question about whose roadmap you are trusting with your infrastructure.
Portkey processes trillions of tokens per month. It will continue to do so. The multi-provider access, the routing, the failover logic: none of that disappears. What changed is who decides where the product goes next.
Palo Alto frames autonomous agents as a new, unmanaged attack surface. Portkey's CEO described the combined goal as establishing "the AI Gateway as the foundational layer of the secure AI enterprise." The product roadmap now answers to a security thesis.
That is coherent strategy for Palo Alto. For teams that chose Portkey for cost routing, developer velocity, or compliance portability, the question is whether a security company will continue prioritizing those capabilities with the same intensity.
Acquisition aside, the two gateways occupy different positions on almost every axis that matters in production.
| Capability | Portkey | LiteLLM |
|---|---|---|
| Throughput (2 vCPU) | 100,000 rpm | 4,800 rpm |
| Deployment modes | Cloud, managed, self-hosted | Self-hosted only |
| Compliance certs | SOC 2 Type II, ISO 27001 | In progress |
| Prompt management | Built-in | No |
| Observability | Built-in | Basic (Prometheus metrics, logs) |
| Guardrails | Built-in | With Enterprise license |
| Fine-tuning | Built-in | No |
| Model access | 200+ LLMs | 100+ LLMs |
| Target uptime | 99.99% | Depends on your ops |
The throughput gap is the number most teams underestimate. A 20x difference on identical hardware means LiteLLM deployments scale horizontally much sooner, and horizontal scaling is where operational complexity compounds.
Portkey ships observability, guardrails, and prompt management as built-in platform features. LiteLLM is thinner: Prometheus metrics and guardrails come with the Enterprise Basic license, not the free proxy, and prompt management is missing entirely. You pay for the license, bolt on your own stack, or go without. For a deeper breakdown of how each stacks up against independent alternatives, see the Portkey and LiteLLM comparison pages.
Open-source does not mean free. LiteLLM's core proxy costs nothing to self-host. The enterprise tier starts around $250/month for Enterprise Basic and scales to roughly $30,000/year for Enterprise Premium.
But infrastructure is the bigger line item. A production deployment handling moderate traffic typically runs $200 to $500 per month in cloud infrastructure alone, before you count DevOps time.
Then there is the operational surface you inherit. LiteLLM requires your team to own deployment, scaling, monitoring, security hardening, and compliance. Your security posture depends entirely on how you harden and operate it. Governance, logging retention, access controls: all yours to build and maintain.
A realistic total for a team running LiteLLM seriously in production is far above the sticker. TrueFoundry's pricing guide puts real-world production TCO at $2,000 to $3,500 per month once the enterprise license, infrastructure, and DevOps time are counted, and the DevOps line dominates: $1,500 to $2,000 of that figure is labor. That is not expensive for what you get, but it is a different proposition than "free." Teams comparing costs against managed gateways (see OpenRouter vs LiteLLM for another angle) need to price the full stack, not just the license.
Palo Alto is a security company. That is their core thesis, their platform strategy, and the lens through which they will orient Portkey going forward.
The problem is that AI gateways serve six distinct buyer personas: security teams want visibility and policy enforcement, finance teams want cost control and model routing, developers want latency and capability-based routing, platform teams want connective tissue across internal AI systems, and governance and privacy teams each have their own requirements. Security is one of six.
Gartner is publishing its first Magic Quadrant for AI Gateways in 2026, formally establishing it as a major enterprise category. Their explicit guidance on the acquisition: "Organizations must manage their AI gateway strategy independent of the AI gateways embedded within larger platforms, due to the likely limitations of embedded AI gateways."
That is a direct statement about the risk of coupling your gateway to a platform vendor's strategic priorities. When one buyer persona (security) controls the roadmap, the other five get what security leaves on the table.
Portkey still fits if your organization already runs Prisma AIRS or plans to, and your primary gateway requirement is security enforcement and governance. The throughput, the certifications, the built-in observability: all still real advantages. The acquisition adds security depth that standalone Portkey never had. For teams inside the Palo Alto ecosystem, the integration will likely get better, not worse.
LiteLLM still fits if your team has strong DevOps capability and needs maximum control over where and how the gateway runs, particularly for compliance requirements that demand self-hosting in your own environment. You trade convenience for control, and for some regulated industries that trade-off is non-negotiable.
Neither fits cleanly if you need a managed gateway that is not owned by a security vendor and is not an ops burden you run yourself. The acquisition left a gap in the middle: production-grade, managed, independent. That is where teams should be evaluating what a neutral AI gateway looks like when neither option's incentives align with yours.
Three questions to pressure-test your current setup:
The portkey vs litellm comparison used to be about features. After April 30, 2026, it is about alignment.
shim provides PII placeholder processing and accounting for native OpenAI, Anthropic, and Google routes. It does not offer semantic caching or cross-provider routing. Self-hosted containers are available only through project-based engagements, not as a packaged self-service product.