How is this cheaper than calling providers directly?

It isn't, on the token itself: every token is resold at the provider's own list price, with nothing added. The saving, when there is one, comes from the router sending a request to a cheaper pool member instead of your most expensive one, whenever a verification check will let that cheaper answer through.

What stops 'cheap' from meaning 'unchecked'?

The verification ladder that runs after every generation, regardless of which pool member answered. A cheap model's output has to clear the same checks a criticality score assigned to that request, and a failure escalates to the pool's strongest remaining model rather than shipping.

Is there a subscription or a cheaper tier I'm missing?

No. There is one fee: a flat rate charged once, when you buy credits. No plans, no seats, no feature gated behind a higher tier.

What exactly do you take a cut of?

Only the initial credit purchase. Tokens, verification calls and triage calls are all billed through at the exact figure the provider charged for that call.

Cheap LLM router

Cheap, without pretending a check didn’t happen.

A credit is one US cent of real upstream spend. Tokens, verification calls and triage calls are all resold at the provider’s own list price, with nothing added on top. The only thing llm11charges for is buying credits in the first place: a flat 5% fee, taken once, at purchase.

No subscription, no per-seat charge, no feature held back for a higher tier. And “cheap” here has a specific meaning: the router sends a request to the least expensive model in your pool that a verification check will actually let through, not the least expensive model regardless of whether it coped.

Start free See full pricing

Buy credits

5% fee, once, at purchase

of credit

$5.25 charged

($0.25 fee)

$25

of credit

$26.25 charged

($1.25 fee)

$100

of credit

$105.00 charged

($5.00 fee)

$500

of credit

$525.00 charged

($25.00 fee)

A $25 top-up costs $26.25: $25.00 lands on your balance, and $1.25 is the fee. Every request after that debits your balance at exactly what the model, the check, and the triage call cost, and nothing more.

Start free, $2.00 of credit on signup

What cheap actually looks like on a receipt

Take a short request, something like 200 tokens of prompt and 40 tokens of completion, the kind of call a typical app makes constantly. Directionally, here is what changes and what does not as it moves through the pool.

Line	If the pool’s most expensive model answered	If the cheapest pool member answered instead
Token cost	A few cents, billed at that model’s own list price for those 240 tokens.	A small fraction of a cent, at the cheap model’s list price for the same 240 tokens.
Triage cost	Near zero either way. One typed decision call, run once, before generation, regardless of which pool member it ends up naming.
Verification cost	A small, proportionate add-on: free schema and reference checks always run, and anything heavier is gated on how critical triage judged the request to be, not on which model answered.
If the cheap answer fails a check	One escalation to the pool’s strongest remaining member, re-checked, then it stops. You pay for both attempts, and the receipt says so rather than hiding the retry in the total.

That is the whole mechanic: a request that would cost a few cents on your most expensive pool model can cost a fraction of a cent when a cheaper pool member can answer it and the checks pass, because the savings figure only exists after the answer has already cleared verification. Your own receipt shows the real numbers for your own traffic, priced against your own pool. We are not going to print a savings percentage here and ask you to trust it; it depends entirely on your pool, your prompts, and how often the cheap model actually holds up.

How you narrow to the cheap end

A pool is what triage is allowed to pick from. Send llm11-fast as your pack and every candidate the router can choose already sits at the cheap end of the live catalogue, priced under $0.50 per million blended tokens.

llm11-fast

The cheapest end of the catalogue. Best for high-volume, low-stakes traffic.

llm11-balanced

A working mix of cheap and capable models. The sane default.

llm11-smart

The most capable models available, priced at whatever that costs.

Set the default per project in settings, or send model or llm11_models per request to override it. A single pinned model id works too, and still gets checked on the way out.

Cheap without a check is a gamble

A router that just picks the least expensive model and ships whatever it returns cannot tell you when that model quietly got something wrong. Nothing about the price it charged you tells you that.

Verification is what turns it into a decision

Every answer, from any pool member, runs through the same ladder: schema and reference checks always, groundedness when you supplied context, heavier checks when criticality earns them.

A failure has a defined outcome

It escalates once, to the strongest remaining model in the pool, gets re-checked, and stops there. The receipt names which check failed and why, so a downgrade is auditable rather than assumed.

Cheap, without pretending a check didn’t happen.

Buy credits

What cheap actually looks like on a receipt

How you narrow to the cheap end

Cheap without a check is a gamble

Verification is what turns it into a decision

A failure has a defined outcome

Go deeper