Sales teams are using language models for more of the routine work: drafting first-touch emails, summarizing a prospect's website, tagging replies, and turning raw lead lists into short research notes. Each of these steps is cheap on its own. The trouble starts when a campaign touches ten thousand leads and every record triggers three or four model calls. This article looks at where the money goes in outbound automation and how to keep it visible and under control.
Where the cost comes from
It helps to map a typical workflow before choosing tools.
Lead enrichment turns a company name and domain into a short description, an industry label and a guess at the buyer's likely priorities. Personalization takes that note and produces an opening line or a full draft. Reply handling reads incoming responses and classifies them as interested, not now, wrong person or unsubscribe. Summaries feed the CRM so that the next person on the account has context.
Costs scale with three variables: how many records you process, how many tokens each call uses, and which model you pick. Records are fixed by the campaign. Tokens are partly in your control through prompt length and output limits. Model choice is the biggest lever, because prices vary widely between models and the cheapest one that meets your quality bar is usually the right one for each step.
Match the model to the task
A common mistake is to use one strong model for everything because it is the one that worked in the first test. A better approach is to treat each step separately.
Classification of replies and enrichment tags rarely need a top-tier model. A smaller, cheaper model with a clear prompt and structured output is usually enough. Personalized drafts deserve a closer look: test two or three models on a sample of real leads, have a salesperson rate the results blind, and pick the cheapest option that the team would actually send.
Put the choice in configuration. If the model name lives in one place, you can change it when a better or cheaper option appears without touching the workflow itself.
Make cost visible per request
Budget control starts with measurement. For each model call, you want to know the model used, the input and output tokens, the latency and the cost, and you want to be able to group those by campaign or workflow step.
Gateways that sit in front of several providers can help here because they put usage and billing in one place. When you compare them, check what the documentation says about transparency. OpenRouter's FAQ, for instance, says that it passes through provider pricing without markup and charges a 5.5% fee (with a $0.80 minimum) on credit purchases, and that it may expire unused credits after one year. Those are clear terms, and they are worth including in your cost model. Read the equivalent terms for any gateway you consider, including top-up fees, minimums and expiry.
Per-call detail matters as much as the price list. ZenMux documents that every call is metered and that per-call cost details and a rate breakdown appear in its log details, with a cost statistics page for the current day or month. Its billing documentation lists separate items for prompt tokens, completion tokens, cache reads and writes, and internal reasoning, which is useful when a model's output includes reasoning tokens that you did not expect to pay for.
Compare routing gateways on cost transparency
If you are weighing cheaper OpenRouter competitors, the headline price is only part of the picture. A practical comparison for a sales automation team covers a short list of questions:
What fees apply when you add funds, and is there a minimum? Do credits expire? Can you see the cost of an individual request after the fact? Can you filter spend by model, key or project? How does billing behave for cached input and reasoning tokens? What happens when a provider fails, and who pays for the retry?
Run the same sample of leads through each option and compare the invoices with your own token counts. A few hundred records is enough to reveal differences in how usage is reported.
| Cost question | OpenRouter | ZenMux | LiteLLM proxy |
| Markup on model prices | Provider prices passed through without markup on inference | Token-level billing; markup terms not listed on the page checked | Spend is calculated from a model cost map |
| Fee when adding funds | 5.5% with a $0.80 minimum on credit purchases; 5% for crypto | Top-ups from $5 to $25,000; fee not listed | Not listed |
| Credit expiry | Unused credits may expire after one year | Expired bonus credits appear in the balance breakdown; other terms not listed | Not listed |
| Per-request cost visibility | Activity page filterable by model, provider and API key; a credits API reports the balance | Request logs, cost analytics and billing statistics | Cost returned in the x-litellm-response-cost response header, plus spend logs |
| Spend controls | Not listed | Per-key credit limit with an email alert threshold, plus RPM and TPM limits | Spend tracked per key, user and team |
Prices and specs as listed on official pages in November 2026; check the live pages before deciding.
Plan for failures without overspending
Outbound automation often runs overnight or on a schedule, so it needs a defined failure path. A fallback model lets a request move to a backup if the first choice is unavailable or rate limited. ZenMux's documentation describes this as a per-request option or a global default set in its console, and recommends choosing a stable, cost-effective model as the backup instead of an expensive flagship, which keeps emergency costs down. It also notes that only one fallback model is supported and that a failure of that fallback returns an error, so your workflow should handle that case by queuing the record for a retry.
Whatever you use, log which model actually answered. A silent switch to a more expensive model can change the economics of a campaign without anyone noticing.
Set guardrails before you scale
A few simple controls prevent most surprises. Cap output length for each step, since drafts rarely need more than a few hundred tokens. Test on a small batch and multiply by the campaign size before launching. Keep a prepaid balance at a level that matches the campaign budget, which both ZenMux's Pay As You Go plan and OpenRouter's credit system support in different ways. Review spend by workflow step weekly, not only at invoice time. And keep a human review step for any message that will be sent under a named person's signature.
A note on fit
ZenMux is a hosted service. Teams that need a fully self-hosted or on-premises gateway for compliance reasons should look at self-managed options instead. Sales data often includes personal information, so check what you are sending to a model and whether your privacy policy and local regulations allow it.
Frequently Asked Questions
How does OpenRouter pricing work?
According to OpenRouter's FAQ, you buy credits and the cost of each request is deducted from them. Prices are shown per million tokens for each model, often with different rates for prompt and completion tokens, and some models charge per request, for images or for reasoning tokens. OpenRouter passes through provider prices without markup but charges a 5.5% fee, with a $0.80 minimum, when you purchase credits.
How does OpenRouter make money?
OpenRouter says it charges a fee when you purchase credits and does not mark up the prices of the underlying providers. The fee is 5.5% with a $0.80 minimum, and 5% for crypto payments. For bring-your-own-key use, its FAQ describes a monthly free allowance on pay-as-you-go accounts and a 5% fee on usage above that allowance.
How much can I use OpenRouter for free?
OpenRouter's FAQ says new users receive a small free allowance and that many models are available at zero price. Those free models are limited to 50 requests per day in total, rising to 1,000 per day once you have purchased at least 10 credits. The FAQ also notes they are usually not suitable for production use, so a paid model is the safer choice for a live sales workflow.
What are some alternatives to OpenRouter that are cheaper?
Cheaper depends on the fee model as well as the model prices. Compare fees on adding funds, credit expiry rules and per-request cost visibility. Options include other hosted gateways and self-hosted software such as LiteLLM, whose documentation describes a self-hosted server with spend tracking. Run the same sample workload through each candidate and compare the invoices with your own token counts.
Conclusion
LLM costs in sales automation are manageable when you treat them like any other operating expense: map the workflow, choose the model per step, measure per request, and read the billing terms of the gateway you use. Comparing options on fees, expiry, per-call visibility and failure behavior will tell you more than a price list will, and a small trial with your own leads will settle most of the questions.