Key Takeaways
- LiteLLM’s open-source proxy is free at any volume — 100+ provider APIs behind one OpenAI-compatible endpoint, with spend tracking, budgets, and rate limiting in the free tier.
- OpenRouter charges no subscription but takes ~5.5% on Stripe credit purchases ($0.80 minimum), 5% via crypto, plus a 5% BYOK fee above $25,000/month (pay-as-you-go) or $200,000/month (enterprise).
- Portkey was acquired by Palo Alto Networks (completed 29 May 2026) and is being folded into Prisma AIRS — treat it as an enterprise security purchase now.
- LiteLLM’s GitHub repo carries ~57.5k stars and ~11.0k forks as of August 2026 — the clearest public adoption signal of the three.
- OpenRouter lists 380+ models and is the fastest path from signup to a working multi-model call.
- The honest crossover: self-hosting is cheaper the moment 5.5% of your monthly model spend exceeds what you pay to run the proxy.
- Skip a gateway entirely if you call one provider at low, predictable volume with no second provider planned.
🏆 Best Choice: Should You Self-Host or Rent the Gateway?
Run LiteLLM yourself if you have somewhere to run it; use OpenRouter if you do not. That single question decides this purchase more reliably than any feature checklist, because all three tools already cover the core gateway job — one API, fallback, cost tracking.
Self-host or hosted gateway?
Use a local LLM if:
- •You already operate services and can patch a proxy — LiteLLM adds no per-request markup at any volume
- •Monthly model spend is high enough that a 5%+ fee is real money (above roughly $500/month it exceeds most small-instance hosting bills)
- •You need to control exactly which provider serves a failed request — e.g. staying inside an EU data-residency boundary
- •You need air-gapped deployment (LiteLLM Enterprise supports this explicitly)
Use a cloud model if:
- •You have no DevOps capacity and do not want to own a running service
- •You want to be calling 380+ models within the hour — OpenRouter needs an API key and nothing else
- •Your spend is modest enough that ~5.5% on funded credits is a rounding error next to the time saved
- •You are already buying Palo Alto Networks security tooling and want the gateway governed inside Prisma AIRS
Quick decision:
- →Free and self-hosted → LiteLLM (Apache-licensed OSS core, ~57.5k stars)
- →Zero infrastructure, models today → OpenRouter (380+ models, ~5.5% credit fee)
- →Enterprise AI security governance → Portkey / Prisma AIRS ($49/mo Production, custom Enterprise)
- →One provider, low volume → no gateway — call the SDK directly
What Is an LLM API Gateway?
An LLM API gateway is a routing layer between your application and multiple model providers that exposes one consistent API. Instead of writing separate integration code for OpenAI, Anthropic, Google, and everyone else, you call the gateway once and it handles provider-specific formatting, authentication, fallback, and — in most implementations — cost tracking and rate limiting.
The payoff only exists once you have more than one provider in production. For a single-provider app the gateway is an abstraction layer with no benefit yet. Use a gateway if provider sprawl is already costing you glue code and on-call time; skip it if it is not.
📍 In One Sentence
An LLM API gateway routes calls from your application to multiple model providers through one consistent, OpenAI-compatible interface, adding fallback, cost tracking, and rate limiting on top of the raw provider APIs.
💬 In Plain Terms
Think of it as a power adapter for AI models. Your code plugs into one socket; the adapter deals with the fact that every provider has a differently shaped plug. If you only ever use one plug, you do not need the adapter.
LiteLLM vs OpenRouter vs Portkey: Full Comparison
No single gateway wins every category: LiteLLM wins on cost and control, OpenRouter on setup speed, Prisma AIRS on enterprise governance. Prices and program details verified against each vendor’s own pages on 27 August 2026 — recheck if this page has been live more than 90 days, since gateway pricing has moved repeatedly this year.
| Factor | LiteLLM | OpenRouter | Portkey (Prisma AIRS) |
|---|---|---|---|
| Best for | Self-hosted teams wanting full control, no fees | Fast multi-model work with no infrastructure | Enterprises already buying Palo Alto security tooling |
| Pricing | Free (OSS core); custom annual (Enterprise) | No subscription; ~5.5% Stripe credit fee, 5% crypto | Free 10k logs/mo; $49/mo Production; custom Enterprise |
| Self-hosting | Yes — core design, plus a managed proxy option | No — hosted-only API access | Private cloud deployment on Enterprise |
| Providers / models | 100+ provider APIs | 380+ models listed | Gateway repo previously advertised 1,600+ LLMs |
| Fallback / failover | Yes — configurable chains in the OSS tier | Yes — automatic, platform-level | Yes — fallbacks and load balancing on all tiers |
| Cost tracking | Yes — spend tracking and budgets in OSS | Yes — per-request credit dashboard | Yes — cost attribution, virtual-key budgets |
| Rate limiting | Yes — OSS tier | Global, account-level (not self-configurable) | Granular controls on Enterprise |
| Affiliate program | None found | None found | None found (Palo Alto partner program is B2B) |
⚠️Warning: Pricing captured before May 2026 no longer describes Portkey. Confirm current Prisma AIRS terms with Palo Alto Networks before quoting a figure from an older article or screenshot.
The Self-Hosted Track: Why LiteLLM Is the Default
LiteLLM is the community default because the free tier is the whole product, not a trial of it. Spend tracking, budgets, rate limiting, and fallback chains all live in the open-source core — the Enterprise tier adds SSO, SCIM, dedicated support, and air-gapped deployment, not the routing itself.
📍 In One Sentence
LiteLLM’s open-source proxy calls 100+ LLM provider APIs through one OpenAI-compatible interface at no license cost and no per-request markup, at any request volume.
LiteLLM — Best Overall Gateway
Free, self-hosted, ~57.5k GitHub stars — the closest thing to a default
LiteLLM’s core proxy is free and open source, routing 100+ provider APIs through one OpenAI-compatible endpoint. It is the only option here with no percentage markup on model spend, which is what makes it the cheapest choice in absolute dollars at meaningful volume. Its ~57.5k stars and ~11.0k forks (August 2026) are the strongest public maintenance signal of the three tools — relevant when you are betting a production routing layer on a project still being maintained in two years. The cost you do pay is operational: you run, monitor, and upgrade a service.
Pros
- +No per-request or percentage fee — free at any volume
- +Fallback chains, budgets, and rate limits all in the free OSS tier
- +Full control over routing logic and data residency
- +Enterprise tier supports air-gapped deployment explicitly
Cons
- –You own the uptime, patching, and monitoring of a running service
- –Enterprise pricing is custom-quoted — no published figure to budget against
- –Wrong choice for a team with zero DevOps capacity
The Hosted Track: OpenRouter and Prisma AIRS
There are honest, non-marketing reasons to rent the gateway instead of running it. You skip driver-free but still real infrastructure work: uptime, upgrades, secrets management, and the on-call rotation that comes with any service in the request path. A hosted gateway also gets you provider coverage you would otherwise integrate one at a time.
The tradeoff is a percentage fee that scales with spend and less granular control over failover. That is a good trade at modest volume and a bad one at high volume — the cost section below gives the exact crossover arithmetic.
OpenRouter — Best Zero-Setup Gateway
380+ models from one API key, no server to run
OpenRouter is the fastest path from nothing to a working multi-model application: sign up, take an API key, and call 500+ listed models the same day. There is no subscription — the fee sits on funding the account, at roughly 5.5% on Stripe credit purchases ($0.80 minimum) or 5% via crypto. Its BYOK fee of 5% only applies once monthly BYOK-routed spend passes $25,000 (pay-as-you-go) or $200,000 (enterprise tier), so light users pay nothing extra there. Fallback is automatic but platform-level: when a provider errors, OpenRouter can route to an equivalent model without you configuring the chain.
Pros
- +No infrastructure at all — an API key is the entire setup
- +380+ models listed, far more than you would integrate by hand
- +No subscription; light BYOK users stay inside the free allowance
- +Automatic provider fallback with no configuration
Cons
- –The ~5.5% credit fee compounds and gets expensive at high volume
- –Rate limiting is global and account-level, not self-configurable
- –Less control over exactly which provider serves a failed request — a problem under strict data-residency rules
Portkey / Prisma AIRS — Enterprise Security Buy
Now the AI gateway inside Palo Alto Networks’ security platform
Portkey still does the gateway job — automatic fallbacks, load balancing, cost attribution, and virtual-key budgeting across all tiers, with a free Developer plan at 10,000 logs/month and a $49/month Production plan at 100,000 logs/month. What changed is the category: since Palo Alto Networks completed the acquisition on 29 May 2026, it is positioned as the AI gateway inside the Prisma AIRS security platform rather than a standalone router. That is a strong fit if you are already procuring enterprise AI governance and a mismatch if you wanted a nimble independent tool.
Pros
- +Fallbacks and load balancing available on every tier, including free
- +Cost attribution and virtual-key budgeting built in
- +Private cloud deployment available on Enterprise
- +Bundles gateway routing with broader AI security governance
Cons
- –No longer independent — roadmap and pricing now follow a large security vendor
- –Pricing and tier structure captured before May 2026 may already be stale
- –Wrong fit if vendor-neutrality is part of why you wanted a gateway
What Does Each Gateway Cost at Scale?
LiteLLM’s open-source core stays free regardless of volume; OpenRouter’s cost is a percentage of every dollar you fund; Portkey’s scales with log volume and contract terms. The table below is straight arithmetic on OpenRouter’s published ~5.5% Stripe credit fee — it is not an estimate, and it is the number that decides this purchase.
We deliberately do not quote a dollar figure for running LiteLLM: the proxy is a small always-on service, and what that costs depends entirely on where you run it. Put your own hosting bill in the last column and the decision makes itself.
| Monthly model spend | OpenRouter credit fee (~5.5%) | LiteLLM licence cost | Self-host wins below |
|---|---|---|---|
| $200 | ~$11/mo (~$132/yr) | $0 | $11/mo |
| $500 | ~$27.50/mo (~$330/yr) | $0 | $27.50/mo |
| $2,000 | ~$110/mo (~$1,320/yr) | $0 | $110/mo |
| $10,000 | ~$550/mo (~$6,600/yr) | $0 | $550/mo |
Use LiteLLM if monthly model spend is high enough that a 5%+ markup is real money in absolute terms. Use OpenRouter if your spend is modest enough that the fee costs less than the engineering hours you would spend running a proxy. For the same build-vs-rent arithmetic applied to hardware, see the local LLM cost calculator.
•Info: The fee applies to credits you buy, not to each request — and OpenRouter’s separate 5% BYOK fee only starts above $25,000/month of BYOK-routed spend (pay-as-you-go) or $200,000/month (enterprise). Below that allowance, BYOK adds nothing.
Why Portkey’s Ownership Change Matters for This Decision
Portkey is no longer an independent product: Palo Alto Networks completed its acquisition on 29 May 2026 and is folding it into the Prisma AIRS security platform. This is not a rebrand. A gateway adopted today as a lightweight indie tool can be re-priced, re-scoped, or bundled into a larger security suite within a product cycle.
If you were evaluating Portkey specifically because it was a nimble startup product, that premise no longer holds. Re-verify current terms directly with Palo Alto Networks before committing — acquired products routinely change pricing, tier structure, and support models in the first 6–12 months of integration.
📍 In One Sentence
Palo Alto Networks completed its acquisition of Portkey on 29 May 2026 and is integrating it as the AI gateway inside the Prisma AIRS security platform.
How Does Each Gateway Handle Fallback and Failover?
All three support automatic fallback when a provider fails — the difference is who writes the failover policy. LiteLLM and Portkey both let you define fallback chains and load-balancing rules in configuration you control. OpenRouter handles it at the platform level: when a provider errors it can route to an equivalent model automatically, with less granular control over exactly which one.
- LiteLLM — fallback chains defined in your own config; the most control over which backup model serves a failed request
- Portkey / Prisma AIRS — automatic fallbacks and load balancing available on all tiers, configured in the platform
- OpenRouter — automatic, platform-level provider fallback; no chain to configure and no guarantee about which provider takes over
Who Should Use Each Gateway?
Developers with infrastructure should default to LiteLLM; teams that want no-ops multi-model access should use OpenRouter; enterprises inside a Palo Alto stack should evaluate Prisma AIRS. Beyond that split, team size, compliance requirements, and tolerance for operational overhead decide the rest.
- Solo developers and small startups prototyping across models → OpenRouter. One API key, 380+ models, no server to babysit. Skip it if you are cost-sensitive at high volume — the credit fee compounds.
- Platform teams running production multi-model apps → self-hosted LiteLLM. Free, the community default, and you keep control of routing and data residency. Avoid it with no capacity to operate a service.
- Enterprises already procuring Palo Alto Networks products → Portkey inside Prisma AIRS. It now ships as a governance and security layer, so the pitch is different than it was a year ago.
- Regulated industries needing air-gapped deployment → LiteLLM Enterprise supports this explicitly; verify current Prisma AIRS deployment options directly with Palo Alto Networks, since acquisition integrations change what is offered short-term. See also enterprise compliance for local LLMs.
- Teams calling one provider at low volume → no gateway at all. Call the SDK directly and revisit when you add a second provider.
LLM Gateways in the EU, Japan, and China
Where the gateway runs decides where your prompts go. That makes gateway choice a compliance decision in three major markets, not just a cost one.
Common Mistakes When Choosing an LLM Gateway
- 1Assuming "gateway" means the same thing across all three tools
Why it matters: LiteLLM is a self-hosted proxy, OpenRouter is a hosted API, and Portkey is now part of a security platform. Read each vendor’s current product page before comparing feature lists — do not assume parity. - 2Picking OpenRouter for high-volume production without modelling the fee
Why it matters: At $10,000/month of model spend the ~5.5% credit fee is roughly $550/month, or $6,600 a year. Run that arithmetic against your actual spend before committing, not just the convenience factor. - 3Adopting a gateway before you have a second provider in production
Why it matters: A gateway’s core value is multi-provider routing and fallback. With one model from one provider you are adding an abstraction layer, a new failure mode, and a learning curve for no payoff yet. - 4Quoting Portkey pricing captured before May 2026
Why it matters: Pricing pages and blog posts from before the Palo Alto acquisition may no longer reflect current Prisma AIRS terms. Confirm directly with the vendor before basing a procurement decision on an old screenshot. - 5Treating GitHub stars as a feature comparison
Why it matters: LiteLLM’s ~57.5k stars indicate community trust and the likelihood the project is still maintained in two years — not that it has more features than Portkey or OpenRouter. Check the comparison table, not the adoption signal.
Skip a Gateway Entirely If…
If you call a single LLM provider at low, predictable volume with no near-term plan to add a second, skip all three. A gateway adds operational complexity (self-hosted) or a percentage fee (hosted) in exchange for a multi-provider fallback benefit you are not using yet.
Call the provider’s SDK directly and revisit this decision the day you actually add a second model or provider. The same logic applies one layer down: if you are still deciding whether to call hosted APIs at all, read local LLMs vs cloud APIs first — the cheapest gateway is the one you never needed.
💡Tip: A useful trigger: add a gateway the first time you write provider-specific branching in your own code. That branch is the glue code a gateway exists to delete.
Frequently Asked Questions
What is an LLM API gateway?
It is a routing layer that gives your application one consistent API for calling multiple LLM providers, typically adding fallback, cost tracking, and rate limiting on top of the raw provider APIs. You call the gateway; it handles provider-specific formatting and authentication. The benefit only materialises once you have more than one provider in production.
Is LiteLLM really free?
The open-source core is free and self-hosted with no licence fee, per LiteLLM’s own pricing page as of August 2026. There is no per-request markup at any volume. The Enterprise tier — SSO, SCIM, dedicated support, air-gapped deployment — uses custom annual pricing sized to your request volume and includes a 30-day trial.
Does OpenRouter charge per request?
Not directly. OpenRouter’s fee is on funding your account: roughly 5.5% on Stripe credit purchases with a $0.80 minimum, or 5% via crypto. A separate 5% BYOK fee applies only once monthly BYOK-routed spend exceeds $25,000 on pay-as-you-go or $200,000 on the enterprise tier.
Is Portkey still an independent company?
No. Palo Alto Networks completed its acquisition of Portkey on 29 May 2026 and is integrating it as the core AI gateway inside its Prisma AIRS security platform, per Palo Alto Networks’ own press release. Treat any pre-May-2026 pricing or roadmap information as potentially stale.
Which gateway supports the most providers?
OpenRouter lists 380+ models on its own site as of August 2026. LiteLLM’s documentation describes 100+ provider integrations — a different unit, since one provider exposes many models. Portkey’s gateway repository has previously advertised 1,600+ LLMs, though that figure should be re-verified given the acquisition.
Can I self-host Portkey?
Portkey/Prisma AIRS offered private cloud deployment at the Enterprise tier per pre-acquisition pricing pages. Confirm current self-hosting terms directly with Palo Alto Networks, since acquisition integrations frequently change deployment options in the first year.
Do any of these three have an affiliate program?
No checkable public affiliate or referral program was found for LiteLLM, Portkey, or OpenRouter as of August 2026. Every product link on this page is a plain, disclosed product link that earns PromptQuorum nothing. If that changes, this page will be updated with disclosure language matching the actual program terms.
How do I decide between self-hosting and a hosted gateway?
Compare 5.5% of your monthly model spend against what it costs you to run a small always-on service. Below that crossover, OpenRouter is cheaper once your time is counted; above it, self-hosted LiteLLM wins on absolute dollars. Add a second test for compliance: if you must be able to state which provider handled a request, self-host regardless of the arithmetic.
Final Verdict
- Use LiteLLM if you can self-host and want zero recurring fees with full routing control — next step: deploy the open-source proxy and point your app’s base URL at it.
- Use OpenRouter if you want multi-model access today with no infrastructure — next step: create an API key and fund the account, budgeting for the ~5.5% credit fee.
- Use Portkey / Prisma AIRS if you are already buying Palo Alto Networks security tooling and want gateway routing bundled in — next step: contact Palo Alto Networks to re-verify current post-acquisition terms.
- Skip all three if you call a single provider at low volume with no near-term multi-provider plans — next step: revisit when you add a second provider.