Guides

Guide

Token Meter, OpenRouter and LiteLLM: how savings are measured, and what each costs

Updated

All three can send your AI calls to cheaper models or reuse cached work. They differ in how they tell you what that saved, and in what they charge. This page compares both, as of 20 November 2026, from each product's public pages and our own code. It makes no claim about how much any of them will save you.

The short answer

  • Measurement. OpenRouter shows what prompt caching saved, not what routing saved. LiteLLM measures its Auto Router against one model, the most expensive in the hardest tier of your config, net of its classifier's cost. It leaves out turns with no cache history, and its totals add $0 for a request it can't price. Token Meter measures each auto request against the priciest model auto considered for it that your rules allowed, and does not deduct its classifier's cost or the prompt cache a switch loses. Token Meter also shows the saving against a baseline you choose, and lists the requests it couldn't price instead of counting them as $0, in its caching and batch figures too.
  • Fees. None of the three marks up tokens. OpenRouter's Standard credit fee is lower than Token Meter's: on $100.00 of credit, $5.50 against $8.00. LiteLLM's open-source version charges nothing, and you run it yourself.

How does each one report what it saved you?

A routing saving compares a model that ran with one that didn't, so the figure depends on which model it is compared with and what is assumed about that model. These are the choices each product makes.

How each product reports savings
WhatToken MeterOpenRouterLiteLLM
Saving from routing to a cheaper modelYes, in dollars. Worked out for each request, then totalled overall, by model pair and by team on Routing activity (the Activity link on Router › Overview). It is an estimate: nobody sent those requests to the other model.No figure against another model. Its Jev Router can return the candidate models' estimated costs, which it calls “not final billed costs”.Yes, for its Auto Router (in beta), in daily tables and on its Cost Optimization page.
What the saving is measured againstThe headline: for a request sent to auto, the priciest of the models auto considered for it that your rules allowed. It is not checked against what the request needed, such as context size, tools or a cost cap. A tiered request is measured against the premium model you choose in Router defaults, and stays unpriced until you choose one. A model you named in the request is measured against itself. Beside it: auto and tiered requests against your Router defaults model.Not applicable.The most expensive model in the hardest tier of your Auto Router config. The saving is net of the cost of the Auto Router's own classifier.
Prompt caching the baseline would have hadNot modelled. The other model is priced with the request's own cache reads and writes, so a switch that cost the conversation its cache can look like a bigger saving than it was.Not applicable.Modelled for Anthropic's native /v1/messages only. Turns with no cache history are left out of both sides.
A model with no known priceThe routing saving leaves the request out, counts it and lists the reason. The caching and batch figures leave it out too and show how many requests weren't priced. With nothing priced, they show no figure rather than $0.Not applicable.Its daily totals add $0 for a request it can't price. Its views show how many requests were compared.
Provider prompt cachingIn dollars on Router › Overview, for all your tracked traffic: cache reads at their own rate, net of what cache writes cost extra. Can be negative.cache_discount on each request, and usage_cache in total through its Analytics API. Negative for cache writes.A prompt caching bucket in its daily savings: the read discount minus the write premium.
Batch discountsIn dollars on Router › Overview, against the same call made directly.Batch models are billed at the provider's batch price. We found no separate saved figure.Not one of its four savings buckets.
Answers reused from a response cacheEstimated in dollars: each hit priced as the call it replaced. Hits that can't be priced are counted apart.Billed at $0 and marked in Activity. We found no figure for what the hit avoided.Logged with $0 spend. Each hit gets an estimated saving, sent to OpenTelemetry but not shown in its dashboard. A request to show it is open on GitHub.
Several kinds of savingShown side by side, not added into one total. When the router sends a call to a batch, that discount counts once, in the routing saving. The batch figure covers the batches you send yourself.Prompt caching, plus a general discounts total in its Analytics API.Four buckets: compression, prompt caching, Auto Router, and the part of prompt caching the gateway set up.

The baseline matters most. Compared with a very expensive model, almost any routing looks like a big saving. LiteLLM picks one such model from your config. Token Meter's headline picks one for each request, as the table shows. Compared with the model you would really have used, the figure is smaller and closer to money you wouldn't otherwise have spent, which is why Token Meter shows that one beside it. Neither is a measured comparison, and neither says anything about answer quality. Token Meter's routing savings methodology sets out its rules in full, including what it leaves out.

What does each one charge?

What each product charges
WhatToken MeterOpenRouterLiteLLM
Markup on tokensNone. Router usage is charged at the provider's list price.None. Models are billed at the provider's price.None. You pay your providers directly.
Fee when you buy credit8% of the credit, at least $1.50, charged once when you buy. Credit charging is rolling out workspace by workspace.Standard plan 5.5%, at least $0.80. Business plan 8%. Crypto 5%. Enterprise: discounts available.No credit. It is software you run.
Calls on your own provider keysNo fee on the tokens: you pay the provider directly. Each call counts as an event toward your plan's allowance, and Growth and Scale charge $14 per million events above it. Once credit charging is on for a workspace, routed calls (Claude Code aside) run on Token Meter's keys and are paid from credit.Free up to $25,000 a month of list-price usage on Standard and Business, then 5%. Not on the Free plan.No fee in the open-source version.
SubscriptionFree $0, Growth $49 a month, Scale $199 a month, Enterprise custom. Each plan includes a number of routed requests a month. Growth and Scale include $25 and $75 of router credit a month, with no fee.None for pay-as-you-go. Enterprise is priced by contract.Open source (MIT, apart from an enterprise folder under its own licence) is free. Enterprise is a paid annual licence, priced by sales, and its Auto Router is listed as an add-on.
Who runs itToken Meter, hosted.OpenRouter, hosted.You: your own servers, plus Postgres for budgets and spend.

The credit fee on a few purchase sizes, before tax:

Credit fee by purchase size
Credit boughtToken Meter feeOpenRouter Standard fee
$20.00$1.60$1.10
$100.00$8.00$5.50
$500.00$40.00$27.50

Token Meter's fee is higher, and Token Meter's paid plans have a monthly price where OpenRouter's pay-as-you-go has none. The credit included with Growth and Scale carries no fee. See Token Meter pricing for what each plan includes.

Which one fits?

  • OpenRouter if you want one key for many models with no subscription and the lower credit fee.
  • LiteLLM if you want open source you host yourself, with no fees for the open-source version, and you are ready to run it.
  • Token Meter if you need to show what routing saved on your own traffic, against the priciest model and against a baseline you set, with the requests it couldn't price listed rather than counted as zero, and what caching and batch saved beside it.

They also work together. Token Meter can track OpenRouter spend by customer and team, and its router can send calls through your own LiteLLM proxy.

What this page doesn't claim

No savings percentage

We don't quote a percentage saved, for Token Meter or anyone else. A saving depends on your traffic, your prompts and the model it is measured against, and we have no customer results to publish yet.
  • OpenRouter and LiteLLM facts are as their public pages and code read on 20 November 2026. Both change often. Where a row says “we found no”, we looked in their docs and didn't find one; it may exist elsewhere.
  • We didn't sign up for, or call, either product to write this page.
  • This page compares measurement and fees only, not model choice, speed, reliability or answer quality.

Sources

All read on 20 November 2026.

See your own savings

Send your calls through the Token Meter router, set the model you would otherwise have used, and Routing activity shows what routing saved on your own traffic against the priciest model and against that model, with anything it couldn't price.