Platform

Two halves of the same bill.

Running a model costs what you send it and what it takes to hold and execute it. The industry has spent a decade on the first. R-Sym works on both.

The argument

One side moves. The other never has.

Prompt caching, shorter prompts, cheaper models, better retrieval — every one of these reduces what reaches the model. None of them change what the model costs to run once it gets there.

Widely addressed

What you send

Tokens, context, retrieval. Understood, competitive, and where nearly all the effort has gone.

  • Prompt cachingPROVIDER
  • Context compressionTHIRD PARTY
  • Retrieval and RAGECOSYSTEM
  • Smaller modelsTRADE-OFF
Where we work

What it takes to run

Memory residency, how weights reach the device, and what has to execute at all. Fixed cost, treated as a constant.

  • Device memory requiredFIXED
  • Weights held residentFIXED
  • Layers executed per tokenFIXED
  • Hardware class neededFIXED
Surfaces

Four surfaces, one system.

Reduction and context are deployed and billing today. Trust is what makes both auditable. Runtime is the second half.

01 / Send less

Reduction

Reduction cuts what a provider bills for on a call you were already making. The model receives what it needs to answer correctly, and what it received can be reconstructed exactly — reduction is recoverable, not lossy.

Because the effect lands on the provider's own invoice, you never have to accept our measurement. The number is theirs.

Reduction →
02 / Hold more

Context

Three things that are usually treated separately: fitting more working context into the same window, keeping that context alive across calls and restarts, and addressing one context surface from more than one provider.

Corpus retrieval sits here too — reasoning across a large local corpus without standing up a vector pipeline to do it.

Context →
03 / Run on less

Runtime

The same approach applied a layer down: what stays resident in device memory, how weights get there, and what actually has to execute to produce a token.

In development. We are not publishing figures on this work yet.

Runtime →
04 / Prove it

Trust

Every dispatch produces a signed receipt. Reduction claims, billing, and the reconstruction check are all recorded against a hash chain you can verify independently.

An efficiency claim you cannot audit is a marketing number. This is the part that makes the rest of it a contract.

Trust →
How it composes

One call, four surfaces.

Nothing here is a separate integration. A single dispatch passes through whichever surfaces your account has enabled.

01

Submit

Your existing call, pointed at the endpoint. Your provider key, your provider account.

02

Reduce

Billed input is cut. Context is resolved from whatever is already held.

03

Dispatch

The provider is called normally and bills you normally, for less.

04

Receipt

The result returns with a signed record of what happened and what it saved.

Get started

Point an existing call at it.

Bring your own provider key. Billing stays with your provider. The difference lands on their invoice.