Build beyond one model call.

Reason over entire datasets. Run on any model. Meter every call to the user who made it.

You are on the list. We will email .

Early access · provider rates pass through with 0% ModelRelay fee · $1 credit included

Recursive model call
query: Why did retention fall last quarter?
01 · mount external working set2.4M records
02 · inspect and reduce with codePython
03 · send selected slices12 calls · 2 models
04 · assemble and verifyanswer ready
Full working set stays externalselected slices enter calls

One platform · three connected capabilities

01

Reason

Go beyond one prompt.

The data is not the prompt.

A reasoning model thinks inside one context window. A Recursive Language Model (RLM) keeps the complete working set in an external environment, then writes code to inspect and reduce it before sending selected slices for model judgment.

Customer intelligence

Find patterns across years of support tickets, calls, and product events.

Code and operations

Trace behavior across a codebase, logs, incidents, and changing dependencies.

Research and diligence

Cross-reference filings, contracts, transcripts, and source material as one investigation.

10M+

tokens handled in published RLM evaluations

0.04 → 58

Same model, run as an RLM: near-zero to 58/100 on questions that require cross-referencing the whole input · OOLONG-Pairs

Published RLM research

Evidence for the pattern, not a guarantee for every workload.

02

Connect

Any model. One API.

Rank models by what matters to you.

Weight latency, throughput, cost, and intelligence for your workload—then call any model through the same API.

Current benchmarks

Compare models on the same controlled task, with output and performance you can inspect.

Your requirements

Set the weights for latency, throughput, cost, and intelligence.

One-line switch

Ship the winner by changing the model string, without rebuilding around another provider.

From benchmark to production

The model you rank is the model you call. Same request, different string.

Call any model
ModelRelay · POST /api/v1/responses Drop-in · OpenAI /v1/responses · Anthropic /v1/messages
curl https://api.modelrelay.ai/api/v1/responses \
-H "Authorization: Bearer $MODELRELAY_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"claude-sonnet-5","input":"Hello"}'

Change the model. Keep everything else.

Explore models and tune the ranking
03

Monetize

Usage in. Revenue out.

Make every AI call accountable and billable.

Know who made every call, what it cost, and what their plan allows. ModelRelay connects model usage to the end user who initiated it.

Customer identity

Pass the signed-in user so direct calls, root calls, and subcalls resolve to the person who triggered them.

Tiers and budgets

Define plans, model access, credits, and limits without building a second control plane.

Your price, your margin

Set what customers pay while ModelRelay keeps the underlying model cost attributable.

Explore AI monetization
One platform, two deployment tracks

Start hosted. Move the data plane when required.

Hosted API
Default

Start immediately.

ModelRelay operates the gateway, RLM runtime, and usage accounting.

Private / VPC

Bring the runtime to your data.

Run the data connector and execution environment in your infrastructure.

Model-provider boundary: in either mode, the providers you choose receive the instructions, schema, and selected data needed for their calls.