LLM Gateway Pricing in 2026: LiteLLM vs OpenRouter vs Portkey, What You Actually Pay

IT Solutions
Abstract glowing network hub routing luminous streams between server nodes in a dark modern data center
November 8, 2026 Updated Nov 16, 2026 9 min read 81 views

LLM Gateway Pricing in 2026: LiteLLM vs OpenRouter vs Portkey, What You Actually Pay

Once your application talks to more than one model provider, an LLM gateway stops being optional: it is the single endpoint that routes requests, holds your API keys, enforces budgets and logs spend. But the pricing pages do not compare cleanly. One vendor charges a percentage when you buy credits, another charges a flat subscription for log storage, and a third gives you the software for free and quietly moves the bill to your infrastructure and on-call rotation. This guide breaks down what the three mainstream options. LiteLLM, OpenRouter and Portkey, actually cost in 2026, with every figure taken from the vendors' own pricing pages, and gives you the decision rules by team size and monthly spend.

What an LLM Gateway Does (and Why You Need One)

An LLM gateway sits between your application and model providers. OpenAI, Anthropic, Google and dozens more, and exposes one OpenAI-compatible API. The layer handles four jobs that every team eventually rebuilds by hand anyway:

  • Key custody and virtual keys: provider keys live in the gateway, and each app, team or customer gets a virtual key with its own budget and rate limit.
  • Routing and fallback: when a provider errors or rate-limits, requests fail over to another provider or model automatically.
  • Spend tracking and budgets: per-request cost attribution, hard spend caps, and the monthly bill nobody discovers late.
  • Observability: request logs, latency traces and audit history: the compliance artifact that shows who prompted what.

Skipping the gateway does not eliminate these functions; it distributes them through your codebase as ad-hoc retry logic and a shared API key in a .env file. The only real question is whether you buy the layer, host the layer, or inherit it on credits.

The Three Pricing Shapes

Every gateway on the market bills in one of three shapes, and the shapes behave completely differently as your volume grows:

  • Percentage on credits (OpenRouter): model inference passes through at each provider's list price with no markup; the platform fee is charged when you top up credits. 5.5% on the Standard plan and 8% on Business, with a $0.80 minimum, per OpenRouter's pricing page (checked November 2026). Bring-your-own-key traffic on Standard and Business includes $25,000 per month of list-price inference with no fee, then 5% above that.
  • Flat subscription on logs (Portkey): no token markup at all. The free Developer tier records 10,000 logs per month with 3-day retention and is explicitly not for production; Production is $49 per month for 100,000 logs with 30-day retention, guardrails and RBAC, then $9 per additional 100,000 requests up to 3 million, per Portkey's pricing page (checked November 2026). The gateway itself is open source and free to self-host.
  • Free software, paid operations (LiteLLM): the MIT-licensed proxy costs $0 in license fees at any spend level. You run the process, the PostgreSQL and Redis backing stores, the version pinning and the pager. Enterprise adds SSO, SCIM, audit logs and support, priced by annual gateway request capacity, quote-only.

A worked example makes the shapes concrete. At $100 per month of inference, OpenRouter's 5.5% credit fee adds $5.50, trivial. Portkey's flat $49 is a 49% surcharge, brutal. LiteLLM adds $0 in fees but costs an engineer's attention. At $10,000 per month the ranking inverts: the same $49 Portkey plan is now 0.49%, OpenRouter's percentage still rides at 5.5% ($550), and LiteLLM stays at $0 software while infrastructure runs roughly $50-$200 per month for a production proxy (estimate per independent gateway-fee audits published in 2026, e.g. Vynaris's normalized gateway-fee table, verified July 16, 2026). The same plan is the most expensive option for a prototype and the cheapest for a serious deployment, nothing else about the choice changes.

LiteLLM: The Self-Host Default

LiteLLM is the most-deployed open-source answer: an OpenAI-compatible proxy covering 100+ providers with load balancing, fallbacks, budgets, virtual keys and Prometheus metrics in the free core. The fastest setup is a one-command pip install of the proxy, a config listing your models, then swapping your client's base URL, a change measured in one line. The catch is operational, not financial. Independent 2026 comparisons (requesty.ai's LiteLLM vs OpenRouter analysis, November 2026) flag two supply-chain incidents this year, malicious versions 1.82.7/1.82.8 on PyPI in March and 1.107.0 in October, which means pinning versions and reading release notes is part of the job. LiteLLM's own docs report about 2 ms median gateway overhead against roughly 100 ms total request latency, so performance is not the concern; ownership is. Pick LiteLLM when keys must stay in your VPC, you have an engineer who will own the proxy, and the $0 license genuinely beats a percentage you would otherwise pay.

OpenRouter: Zero Operations, Percentage on Credits

OpenRouter is a hosted routing network: buy credits, call one endpoint, and it routes across 80+ providers and 600+ models with automatic failover (per its pricing and models pages, checked November 2026). There is no markup on per-token prices: the business runs entirely on the credit-purchase fee (5.5% Standard, 8% Business; the Business tier adds EU or US in-region routing) and 5% on BYOK traffic beyond the $25,000 monthly free allowance. Its reliability record is worth knowing before you standardize on it: the OpenRouter API failed for 38 and 35 minutes on two February 2026 incidents (documented in ai-toolfinder's gateway audit), and the company raised a $113M Series B in May 2026 at a reported ~$1.3B valuation, which is at least evidence of staying power. Pick OpenRouter when engineering time is scarcer than the fee and you want the widest catalog with nothing to operate.

Portkey: The Governance Control Plane

Portkey pairs an open-source gateway (12K+ GitHub stars, universal API, fallbacks, guardrails) with hosted observability: the $49 Production tier is really a logging and governance subscription, metered on recorded logs rather than dollars. The 2026 caveat is ownership: Palo Alto Networks closed its acquisition of Portkey on May 29, 2026, folding it into the Prisma AIRS AI Gateway line, with no public long-term roadmap commitment on the standalone product yet. The gateway is MIT-licensed and self-hostable if you want the features without the subscription or the acquisition risk. Pick Portkey when you want managed guardrails, virtual keys and per-team observability this quarter and can live with flat log-metered pricing.

Which Gateway Pays for Itself at Your Spend

  • Under ~$500/month inference: OpenRouter credits. The percentage is a rounding error and you hire nobody. A self-hosted proxy's setup hours alone exceed the fee.
  • $500-$5,000/month with governance needs: Portkey Production at $49 flat, or LiteLLM OSS if someone on the team already runs infrastructure and you want guardrails in your own VPC.
  • Over ~$5,000-$10,000/month: the percentages stop being abstract. Independent break-even analysis puts self-hosting LiteLLM ahead of a 5.5% credit fee somewhere around the tens-of-thousands range per month depending on how you price ops time; at minimum, negotiate an enterprise fee discount or move BYOK inside the allowance.
  • Regulated or air-gapped data: self-host LiteLLM or Portkey OSS behind your firewall, full stop: no hosted gateway terminates your prompts inside your perimeter.

One line items every team forgets: gateway fees and model tokens are separate bills. A workload can blow through a request-metered log allowance while the token bill stays small, and a "free" OSS gateway still costs the $50-$200 proxy infrastructure plus whoever answers the pager at 2 a.m. when a provider rotates keys.

FAQ

What is the cheapest LLM gateway?

LiteLLM's open-source proxy has a $0 software fee at any volume, which makes it the literal cheapest, but you pay in servers (roughly $50-$200/month for a production deployment), databases and engineering time. The cheapest total cost for most small teams is a hosted option: OpenRouter's 5.5% credit fee under about $1,000 per month of inference costs less than the hours a self-hosted proxy consumes.

Does OpenRouter mark up model prices?

No. Per OpenRouter's pricing page (November 2026), inference is billed at each provider's list price on every plan, and there is no token markup. The platform fee is charged when you purchase credits: 5.5% on Standard, 8% on Business, and BYOK traffic above the $25,000 monthly free allowance carries a 5% fee. Model spend and gateway fees are separate lines on your statement.

Is Portkey worth $49 a month?

It is a log-storage and governance subscription, not a per-token toll. If you need 100,000 logged requests per month with 30-day retention, guardrails and RBAC, $49 is predictable and cheap at scale, at 0.49% of a $10,000 inference bill and under 5% of a $1,000 one. Watch the metering unit: Portkey bills recorded logs, not dollars, so high-volume cheap calls can trigger the $9-per-100k overages even when the token bill is small.

Should I self-host LiteLLM or use OpenRouter?

Self-host LiteLLM when keys and prompts must stay in your infrastructure, you have engineers to own PostgreSQL, Redis, version pinning and the proxy's 2026 supply-chain history. Use OpenRouter when you want zero operations, the widest catalog and failover handled for you, and 5.5% on credits beats your cost of engineering time. You can also run both: LiteLLM's docs list OpenRouter as a provider.

Do LLM gateways add latency?

On the vendors' own numbers, very little: LiteLLM documents about 2 ms median overhead against roughly 100 ms total request time, and Portkey's open-source gateway claims sub-millisecond routing. The practical latency risk is not the proxy itself but routing choices: a failover to a slower provider adds real seconds, which is why provider-preference settings deserve the same review as the fee table.

ArcBeta designs and operates private AI infrastructure for Edmonton businesses, including self-hosted gateways in your VPC, budget-enforced virtual keys and air-gapped model deployments. Ask us about a gateway architecture review. For the bigger build-vs-buy picture behind model serving, our earlier breakdown of self-hosted LLM vs API costs works the arithmetic end to end.