Portkey vs LiteLLM: which AI gateway fits your platform team
Portkey and LiteLLM solve the same problem: one OpenAI-style API in front of many model providers, with fallbacks, retries, spend tracking and access control on top. Both have a free open-source gateway. They part ways on where the control plane lives and how you buy the paid version.
Portkey was acquired by Palo Alto Networks in 2026 and is now sold as Prisma AIRS AI Gateway. On 19 November 2026, portkey.ai/pricing returned a 308 redirect to a Palo Alto Networks contact form, so there is no published paid price. LiteLLM, from BerriAI, lists an Open Source plan at $0 with virtual keys, budgets and teams included, plus a custom-priced Enterprise plan with a 30-day trial license emailed without a sales call.
This comparison covers pricing, governance features, deployment models, operating cost and the questions to settle before you route production traffic through either one. For the wider field, see our AI infrastructure and LLMOps software category.
Portkey
BerriAI-litellm
Verdict
Pick LiteLLM if your platform team can run a production service and wants governed multi-provider access without a license fee. Virtual keys, budgets for keys, users and teams, RPM and TPM limits, load balancing and guardrails are all in the free open-source proxy, which lists 140+ provider integrations. The cost moves to your own infrastructure: Postgres, Redis 7.0 or newer once you run more than one instance, autoscaling, upgrades and supply-chain hygiene after the March 2026 PyPI incident.
Pick Portkey if you want a hosted control plane, prompt management and 40+ pre-built guardrails, and especially if your company already buys Palo Alto Networks security tooling. Its hybrid deployment keeps prompts and responses in your VPC while Portkey hosts the dashboard, and the gateway now connects to Prisma AIRS runtime security and red teaming. The catch today is commercial clarity: with no published price, you need a quote before you can budget for anything beyond the free tiers.
If you only need routing and fallbacks, either free gateway works. The decision comes down to whether you'd rather operate the control plane yourself (LiteLLM) or buy it from a security vendor (Portkey).
- Portkey
- Choose Portkey when you want a hosted control plane with prompt management, guardrails and hybrid VPC deployment, and you're comfortable buying through Palo Alto Networks sales.
- BerriAI-litellm
- Choose LiteLLM when you want virtual keys, team budgets and spend tracking free on your own servers, and you have a platform team to run Postgres, Redis and upgrades.
Choose Portkey when…
Mid-size and large engineering organizations with a platform or security team that owns AI access, particularly companies already using Palo Alto Networks products.
- MIT-licensed open-source gateway that runs locally from one npx command, with Docker, Node.js and Cloudflare Workers deployments documented
- Hybrid deployment keeps the gateway, cache and log store holding prompts and responses inside your VPC while Portkey hosts the dashboard
- Configs define fallbacks on chosen error codes, weighted load balancing, timeouts and up to five retries with exponential backoff
- Prompt management with templates, versioning, a playground and deployment of prompts as API endpoints
- 40+ pre-built guardrails on inputs and outputs, plus custom guardrails
- Docs state ISO 27001 and SOC 2 certification, GDPR and HIPAA compliance and AES-256 encryption at rest
- No published paid pricing: the pricing page redirects to a Palo Alto Networks contact form
- Roadmap is now tied to the Prisma AIRS security suite, which may shift attention away from developer features
- Published numbers conflict: model counts range from 250+ to 3,000+ across pages, and latency claims from under 1ms to 20 to 40ms
- Semantic caching, longer log retention, SSO and private deployment sit outside the open-source gateway
Choose BerriAI-litellm when…
Platform and ML infrastructure teams at companies already running Kubernetes and Postgres that want to issue capped keys to many internal teams across many providers.
- Virtual keys, budgets for keys, users and teams, RPM and TPM limits and spend tracking in the free open-source version
- 140+ provider integrations behind one OpenAI-compatible format, including Bedrock, Azure OpenAI, Vertex AI, vLLM and Nvidia NIM
- Python SDK with one completion() call for teams that don't want to run a proxy at all
- Logging callbacks to Langfuse, Arize Phoenix, LangSmith and OpenTelemetry
- Enterprise page publishes support response targets, starting at 1 hour for Sev 0
- 30-day Enterprise trial license emailed with no sales call, for self-hosted use
- You operate it: Postgres, Redis 7.0+ for multi-instance setups, connection pools and autoscaling are your job
- PyPI versions 1.82.7 and 1.82.8 shipped a credential stealer on 24 March 2026, so pin versions and use signed images
- SSO, SCIM, JWT auth, audit logs and tag-based spend tracking are Enterprise features with no published price
- Defaults need tuning: the production guide changes the request timeout from 6,000 to 600 seconds among more than a dozen settings
At a glance
Portkey is an AI gateway and LLMOps platform. The open-source gateway exposes one OpenAI-style API and forwards calls to the provider you configure, with JSON configs for fallbacks, load balancing, retries, caching and guardrails. The managed product adds request logs with traces, cost and usage by team or credential, a vault for provider credentials behind virtual keys, prompt management, role-based access, SSO and audit logs. Portkey's docs describe a free managed plan of 10,000 requests a month; paid tiers are now quoted by Palo Alto Networks.
LiteLLM is an open-source LLM gateway that comes in two forms: a Python SDK and a self-hosted, OpenAI-compatible proxy that listens on port 4000 in the docs. The proxy issues virtual keys mapped to real provider credentials, enforces budgets and rate limits, tracks spend across providers and ships an admin UI. Everything runs on your servers, with prompts, responses and spend logs in your own Postgres. The Enterprise license adds SSO, SCIM, JWT auth, audit logs, secret-manager integration and support SLAs.
Portkey
- Open-source gateway under MIT, free to self-host
- Managed free plan of 10,000 requests a month, per Portkey's docs
- Hybrid deployment: data plane in your VPC, control plane hosted by Portkey
- Prompt templates, versioning and prompts deployed as API endpoints
- 40+ pre-built guardrails and automatic PII redaction
- Now sold as Palo Alto Networks Prisma AIRS AI Gateway
LiteLLM
- Open Source plan at $0, self-hosted, described as free forever
- 140+ provider integrations in one OpenAI format
- Virtual keys, budgets and teams, load balancing and RPM/TPM limits
- Admin UI for models, keys and spend views
- Helm charts and Terraform modules for AWS and GCP
- Enterprise adds SSO, JWT auth, audit logs and custom SLAs at a custom price
Who each is built for
Portkey is built for organizations that want one governed path for every LLM call and would rather buy the control plane than run it. Its hybrid model answers the most common security objection to a hosted gateway, since request and response bodies stay in your data plane and only aggregated metrics go back to Portkey. Since the acquisition, it fits best at companies that already run Palo Alto Networks tooling, because the gateway now acts as the enforcement point for agent traffic inside Prisma AIRS alongside artifact scanning and red teaming.
LiteLLM is built for platform teams that want to hand every product its own capped keys on day one without a procurement cycle. If you already run Kubernetes and Postgres, the free proxy covers budgets, spend tracking and rate limits that many gateways reserve for paid plans. It's a weaker fit if nobody will own a production service, or if compliance needs SSO and audit logs immediately, since those require the Enterprise license.
Portkey
Best for platform and security teams that want a hosted control plane, prompt management and guardrails with prompts kept in their VPC, bought through Palo Alto Networks.
BerriAI-litellm
Best for platform teams that can run their own infrastructure and want free virtual keys, team budgets and spend tracking across 140+ providers.
Comparison matrix
These rows cover what usually decides a gateway purchase: what's free, where the control plane and logs live, which governance features need a paid license and what you'll operate yourself. Figures come from each vendor's site, docs and GitHub, checked on 19 November 2026.
| Criterion | Portkey | BerriAI-litellm |
|---|---|---|
| Pricing model | Free open-source gateway; managed and enterprise plans quoted by Palo Alto Networks | Free open source; custom-priced Enterprise license |
| Starting price | $0 (open source or managed free plan); paid pricing not published | $0 (Open Source); Enterprise on request |
| Free plan or trial | Managed free plan of 10,000 requests a month per docs | 30-day Enterprise trial license, self-hosted, no sales call |
| Deployment | Self-hosted gateway, hosted SaaS, hybrid VPC data plane; private cloud on enterprise | Self-hosted by default; air-gapped and multi-region control plane on Enterprise |
| API | Yes | Yes |
| Budgets and virtual keys | Virtual keys with budgets in the managed product | Virtual keys and budgets for keys, users and teams in the free proxy |
| SSO and audit logs | SSO with any OIDC provider and org-wide audit logs listed on the homepage; plan placement not published | Enterprise license only |
| Prompt management | Templates, versioning, playground, prompts as API endpoints | Not a headline feature; focus is routing, keys and spend |
| Guardrails | 40+ pre-built checks plus custom guardrails | Content filtering, PII masking and safety checks in the free proxy |
| What you operate | Nothing on SaaS; Kubernetes, Helm, Redis and object storage for hybrid | Postgres, Redis 7.0+ for multi-instance, autoscaling and upgrades |
| Ownership | Palo Alto Networks (acquisition completed May 2026) | BerriAI |
Pricing and packaging
Portkey publishes no current price list. On 19 November 2026, portkey.ai/pricing returned a 308 redirect to the Prisma AIRS AI Gateway contact form on paloaltonetworks.com, and that page showed no plans or prices. Two free options remain. The open-source gateway on GitHub is free under the MIT license, and Portkey's docs state that the managed version offers a free plan with 10,000 requests per month, with paid plans offering more requests and features. Earlier versions of the pricing page listed a self-serve paid tier, but that figure can't be confirmed today, so get a written quote before planning around it.
LiteLLM's site lists two plans as of 19 November 2026. Open Source is $0 and includes 140+ provider integrations, logging to Langfuse, Arize Phoenix, LangSmith and OpenTelemetry, virtual keys, budgets and teams, load balancing with RPM and TPM limits, and LLM guardrails. Enterprise has no listed price; the site offers a request-pricing button and a 30-day trial license, which the enterprise page describes as full Enterprise features, self-hosted, emailed with no sales call.
The real cost of LiteLLM is infrastructure and staff time. Its production guide calls Postgres the primary supported database, asks for Redis 7.0 or newer once you run more than one proxy instance, and suggests starting load tests at 1 vCPU and 4Gi of memory per worker, so an 8-worker deployment starts around 8 vCPU and 32Gi. LiteLLM doesn't bill for any of that, but your cloud provider does.
With Portkey, watch logging volume and retention rather than request counts, and confirm retention periods in the contract. Semantic caching is listed in the GitHub README as hosted and enterprise only, so a self-hosted open-source setup gets simple caching alone. For both gateways, remember the fee sits on top of what you pay model providers: neither replaces your OpenAI, Anthropic or Bedrock bill.
Setup and operations
Portkey's fastest start is the open-source gateway: run it with npx or Docker, point your OpenAI SDK at the gateway URL and pass a config header naming the provider. Python and Node.js SDKs are available. Moving to the managed gateway means creating virtual keys for each provider account and swapping the base URL. Most of the learning curve sits in JSON configs for fallbacks, load-balancing weights, retries, caching and guardrails, so keep one config per use case in version control.
Portkey's hybrid deployment needs platform-team time. The docs require Kubernetes 1.20+, Helm 3.x, Redis for the cache, object storage such as S3, GCS or Azure Blob, outbound HTTPS to the control plane and roughly 1 to 2 CPU cores and 2 to 4 GB of memory per gateway instance. The gateway syncs configs every minute and keeps serving from its local cache if the control plane is unreachable. Portkey's docs say no plan imposes request timeouts, so set client timeouts yourself.
LiteLLM's first test is a uv add litellm for the SDK, or the proxy's Docker image with a config file listing models and credentials. Production is a bigger step: Postgres with a sized connection pool, Redis for multi-instance setups, one Uvicorn worker per Kubernetes pod with scaling by pod count, and a dedicated background-job pod. The docs' example of 3 instances with 4 workers and 10 connections each already reaches 120 database connections.
Two LiteLLM settings need care before launch. The master credential must start with sk-, and the salt that encrypts stored provider credentials must never change after you add models, so store both in a secret manager. After the March 2026 incident, in which PyPI versions 1.82.7 and 1.82.8 carried a credential stealer for about 40 minutes, pin versions and verify the cosign-signed Docker images available from v1.83.0-nightly onward.
Before you choose
Answer these with your platform, security and finance leads before you route production traffic through a gateway.
- Does anyone on your team own production services today, and can they take on Postgres, Redis, autoscaling and upgrades for a gateway?
- Must prompts and responses stay inside your VPC, and is a vendor-hosted dashboard acceptable if only metrics leave?
- Do you need SSO, SCIM and audit logs on day one, and can you budget for them without a published price?
- Does your company already buy Palo Alto Networks security products, and would Prisma AIRS integration help or add weight?
- Do product teams need to change prompts without a code release, or is routing and spend control the main goal?
- How long must you keep request logs for debugging and audits, and how many requests a month will that cover?
Frequently asked questions
Short answers to the questions buyers most often ask when comparing Portkey and LiteLLM.
How much do Portkey and LiteLLM cost?
Both have free open-source gateways. Portkey's docs describe a managed free plan of 10,000 requests a month, but paid pricing is no longer published because portkey.ai/pricing redirects to a Palo Alto Networks sales form as of 19 November 2026. LiteLLM's Open Source plan is $0 and its Enterprise plan is priced on request.
Is Portkey now part of Palo Alto Networks?
Yes. Palo Alto Networks announced the deal in May 2026 and its blog dated 29 May 2026 says the acquisition is complete. The gateway is now sold as Prisma AIRS AI Gateway, and Portkey's docs carry a banner saying so. Palo Alto Networks says it will keep supporting existing Portkey customers.
Which is better for team budgets and spend tracking?
LiteLLM includes virtual keys and budgets for keys, users and teams in its free proxy, with spend tracked across providers. Portkey offers virtual keys with budgets and cost views by team or credential in its managed product. If you want this without a contract and can self-host, LiteLLM is the cheaper path.
Can I keep prompts inside my own network?
Yes, with either. LiteLLM runs entirely on your servers and stores prompts, responses and spend logs in your own Postgres, with air-gapped deployment on Enterprise. Portkey's hybrid deployment runs the data plane, including the log store, in your VPC while Portkey hosts the control plane.
Does LiteLLM have an Enterprise trial?
Yes. LiteLLM's enterprise page offers a 30-day trial license with full Enterprise features for self-hosted use, emailed to you with no sales call. That lets you test SSO, audit logs and access controls before negotiating a contract.
Was LiteLLM affected by a supply-chain attack?
Yes. On 24 March 2026, PyPI versions 1.82.7 and 1.82.8 contained a credential stealer for about 40 minutes before PyPI quarantined them. LiteLLM says the official Docker images and LiteLLM Cloud were not affected, and it shipped v1.83.0 through a rebuilt pipeline. Pin versions and verify signed images.
How many providers do Portkey and LiteLLM support?
LiteLLM lists 140+ providers and about 1,800 models. Portkey's pages cite 1,600+ models, its README heading says 250+ LLMs and the Palo Alto Networks announcement says 3,000+ LLMs, MCP servers and agents. Check the provider list in each vendor's docs for the models you depend on.
Portkey and LiteLLM profiles
Portkey
Our Portkey review covers the pricing change after the Palo Alto Networks acquisition, gateway routing, guardrails, prompt management and the hybrid deployment requirements.
BerriAI-litellm
Our LiteLLM review covers the free and Enterprise plans, virtual keys and budgets, production infrastructure needs and the March 2026 supply-chain incident.
Sources
- Portkey pricing page (308 redirect to Palo Alto Networks contact form)
- Palo Alto Networks AI gateway page (redirect target, no prices shown)
- Portkey docs: what is Portkey (free managed plan)
- Portkey docs: hybrid deployment architecture
- Portkey open-source gateway on GitHub
- Palo Alto Networks blog: Portkey acquisition completed
- LiteLLM plans (litellm.ai)
- LiteLLM Enterprise page
- LiteLLM docs: production best practices
- LiteLLM security update, March 2026
- BerriAI/litellm on GitHub