When talking about model gateways online, a common saying is: OpenRouter does not increase the price, and the model is sold at the official original price, so there is no difference between using it and direct connection.

This sentence is only half correct about token.

The unit price of the token is indeed transferred directly based on the price listed by each supplier. Money is collected at the top-up step: there is a 5.5% handling fee for credit card top-ups, with a minimum of $0.80 for a single transaction. After recharging US$10, the 5.5% rate was originally 0.55, but was raised to 0.80 by the minimum amount. The deducted amount by swiping the card is US$10.80, which is equal to an 8% rate. For people who recharge with small amounts multiple times, the actual rate is much higher than the figure on the promotional page. Top-up with cryptocurrencies is a fixed 5% with no minimum.

LiteLLM on the other side is open source, the code is deployed by yourself, and the gateway itself does not charge any money. This week it just rewrote the /v1/messages interface in Rust, and launched the beta version of the portal for inferring spending based on customer statistics. It seems that one charges tolls and the other is free. The choice is simple.

This is not how it is calculated in actual use.

A table first aligns the caliber

OpenRouter LiteLLM
form Hosting service, one key can call all models Open source code, self-deployed agent
Gateway fee 5.5% on deposits (minimum $0.80), 5% on cryptocurrencies Open source version is free; enterprise version starts at approximately $250 per month for third-party translation
upstream account No, you can also hang your own Key (BYOK) You have to go to each store to open an account and get the Key.
free model 20 times per minute; 50 times per day if the cumulative recharge is less than 10 US dollars, and 1,000 times per day if the total recharge is more than 10 US dollars. None, it depends on the upstream you connect to
To be maintained by oneself almost none Agent process, plus a database that stores keys and budgets, and a cache for multi-instance current limiting

There is something easily overlooked in the free model line. The $10 threshold is calculated based on cumulative recharges, regardless of the current balance: if you recharge $10 once and spend it all, the daily limit of 1,000 times will also be retained. If you just want to use a model with the :free suffix for testing, the most economical way is to recharge $10 in one go, instead of $5 in two installments. The latter requires two minimum handling fees.

The same model name, but later the supplier was changed.

The most criticized thing in the OpenRouter community is not the handling fee.

It's the same model name, but the quality of the answers today is different from yesterday. The reason lies in its routing mechanism: an open source model is often hosted by several inference providers, some running original accuracy, and some running FP8 or even lower quantized versions. OpenRouter will switch between them based on price and availability by default. What you wrote in the request is the same model, but the actual order recipient may be different, and the output will also change accordingly.

There are solutions: you can use the provider parameter in the request to specify or exclude inference providers, filter by quantification accuracy, or request only suppliers that do not retain data. However, these switches are not turned on by default. When most people encounter a sudden format error and clumsy tool calls for the first time, they never think that the supplier has changed.

LiteLLM doesn't have this problem. The request will only be sent to whichever company's Key you receive, and the routing rules are all written by you.

The price is another account.

Set up your own gateway, and replace billing with operation and maintenance

LiteLLM is very light when used as a Python library, just to unify the interface formats of various companies. To use its truly valuable features, such as issuing virtual keys to each colleague, setting budgets according to projects, and automatically intercepting excess quotas, you have to run its proxy service and configure a data inventory key and spending records; open a few more instances to do the load, and the current limit count must be put into the shared cache.

You'll find after a few months that the gateway itself doesn't cost anything, it just costs someone to keep an eye on it.

Performance is also a topic that cannot be avoided. The official figure is that the added delay of each request when deployed natively is usually within 10 milliseconds; while the pressure test issued by competing products says that the tail delay will significantly worsen when more than 500 requests per second are made. The latter was written by a competitor. I don’t fully believe it, but it is a fact that agents written in Python suffer under high concurrency. Rewriting the most popular messages interface first this week also shows that the project side knows it well.

With teams making dozens of requests per second, this is not a problem at all.

Those three hours on March 24

There is another layer of risk in deploying it yourself, which many people do not take into account when choosing: the package itself that you put in it.

On March 24, 2026, the two versions of LiteLLM 1.82.7 and 1.82.8 on PyPI were replaced with backdoor packages, which were released at 10:39 and 10:52 UTC respectively. The attacker first compromised the security scanning tool Trivy used in its CI process and obtained the release credentials from there. 1.82.8 is even more ruthless. It stuffs a .pth file, which will be executed as soon as Python starts. There is no need to import litellm at all. Both versions were quarantined after being online for about three hours, with the last clean version being 1.82.6.

Gateways hold the keys to all your model suppliers. It was poisoned, and what was leaked was not the company's quota.

Afterwards, LiteLLM rotated the maintainer accounts and began to disclose the verification values ​​of each version. This is considered a remedy, but it reminds you that self-hosting does not mean more security. It just moves trust from a hosting provider to your dependency chain. Version locking, checking verification values, and not enabling automatic upgrades in the production environment. If these things cannot be done, the hosting service will be more stable.

Let’s look at the shortcomings on both sides together.

OpenRouter's shortcomings: the actual rate for small recharges is relatively high; routing is opaque by default, and models with the same name may be accepted by inference providers with different accuracy; the BYOK fee-free quota has not been aligned on various pages of the official website by the middle of this year. It is best to go to the price page to confirm before settling the account.

The shortcomings of LiteLLM: it requires maintenance; the Python agent has a performance ceiling under high concurrency; there has been a major accident in the supply chain; the price of the enterprise version is not disclosed, and the numbers circulating on the Internet are mostly quoted from competing products.

If you are an individual developer, spend less than a few dozen dollars on a monthly model, and want to try out various models for a key, the time saved by OpenRouter is far more valuable than the 5.5%. Remember to recharge $10 at a time, and then pin the provider of commonly used models.

If you are a team and want to distribute keys to dozens of people, control budgets by project, and already have direct accounts and discounts in each company, LiteLLM is more suitable. The premise is that someone is willing to take care of it and lock the version.

There is no conflict between the two. Many teams treat OpenRouter as an upstream behind LiteLLM: the commonly used Claude and DeepSeek are directly connected to get discounts, while long-tail models use OpenRouter to take the lead, and statistics and budgeting are done in their own gateways.

Related comparisons: "Strata, Ollama, LM Studio: 8GB video memory with 128K context, what the title does not mention is the 64GB memory", "Gemini 4 Hydrogen and DeepSeek V4.1 Flash 1017: The difference is five times or fifty times, it depends on whether the 100,000 tokens hit the cache"