modelrouting
Concurrency-based inference

Unlimited tokens.
Pay for concurrency.

One flat subscription. Every token you generate is included — the only thing we cap is how many requests run at the same time. Buy in seconds via Telegram, pay in USDT, activate on-chain.

responses · live stream
$ curl -N api.modelrouting.io/v1/responses \ -H "Authorization: Bearer arl_…" -d '{"model":"gpt-6.1-sol","stream":true}' ← 200 text/event-stream data: {"type":"response.output_text.delta","delta":"Concurrency is the..."} ✓ done · 4.6s · 12,832 tokens · ∞ remaining
429? just queue it — Retry-After: 1
∞tokens per key, no quota
0prompts or completions retained
429with Retry-After when streams are busy
$49flat price, one key, all models

Why modelrouting

A thin, fast pass-through to the airelay gateway. We sell concurrency, not volume — everything else is on us.

Fast

Direct routing to the airelay gateway. One hop, zero queues of your own.

Zero data retention

Metadata-only logs. Prompts and completions stream through, never persisted.

No token limit

Quota is unlimited. Generate as much as your streams can carry.

Only concurrency capped

Per-stream limits, per key. That is the entire metering model.

Live in four steps

From first message to your first streamed response — usually under five minutes.

1

Ping the bot

Message @modelrouting_bot on Telegram and run /buy — the bot opens an order and shows you a dedicated deposit address.

2

Pay in USDT

Send USDT on TRC20 (Tron) or BEP-20 (BSC) to the order's derived address. Pick your chain with /paytron or /paybsc.

3

Auto-activation

The moment the transfer confirms on-chain, your arl_… key goes live with its limits. No email, no onboarding.

4

Wire up your harness

Point base_url at api.modelrouting.io/v1 in Codex, OpenAI SDK or curl — done.

Pricing

One price, no tiers, no per-token math. Pay in USDT through the Telegram bot.

Top-up

Add capacity

$49 / per top-up

  • +1 LLM stream, +2 tool streams
  • Stacks on your existing key
  • Extends the subscription 30 days
  • Same unlimited quota
Top up

Paid in USDT (TRC20 · BSC) via the Telegram bot. Every order gets its own derived deposit address — surplus stays there, activation is automatic.

Models

Served live from the airelay catalog — this list updates with the platform. Native /v1/responses endpoint, hosted tools included.

Live catalog loads from /v1/public/models.

Wire up your harness

You get one key — arl_… — that works with any Responses-API-compatible harness. Point base_url at .

# ~/.codex/config.toml
model = "gpt-6.1-sol"
model_providers.modelrouting =
  { name = "modelrouting",
    base_url = "",
    wire_api = "responses",
    env_key = "MR_API_KEY" }

export MR_API_KEY="arl_…your-modelrouting-key"

Every Codex session then runs through modelrouting — unlimited tokens, concurrency-capped.

from openai import OpenAI

client = OpenAI(
    base_url="",
    api_key="arl_…your-modelrouting-key",
)

resp = client.responses.create(
    model="gpt-6.1-sol",
    input="Explain concurrency limits in one line",
    tools=[{"type": "google_search", "google_search": {}}],
)
print(resp.output_text)
curl -N /responses \
  -H "Authorization: Bearer arl_…your-modelrouting-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "input": "What is the weather in Hanoi?",
    "tools": [{"type": "google_search", "google_search": {}}]
  }'

Hosted tools ship with your key: google_search and x_research. Each running tool call occupies one tool stream.

Coming soon. Anthropic /v1/messages support is on the roadmap. Until then, use the native /v1/responses endpoint above with any Responses-compatible harness.

FAQ

Why no token limit?

We route inference through airelay and sell concurrency instead of volume. Your subscription has an unlimited token quota; when your streams are busy you get a 429 with Retry-After: 1 — back off, or buy another stream.

What counts as a stream?

A LLM stream is one in-flight /v1/responses execution. A tool stream is one in-flight hosted-tool execution (search, research). Limits are enforced per key, per request — not per hour.

How does payment work?

The Telegram bot opens a short-lived order and derives a dedicated USDT deposit address for it. Send USDT on TRC20 or BSC; activation is automatic once the transfer confirms. Manual settlement is a fallback for edge cases.

What do you store?

Metadata only: which key, which model, status, duration, byte counts. Prompt and completion content pass straight through to the upstream and are never written to disk.

Can I add more capacity later?

Yes — the /add command adds one bundle of capacity (+1 LLM stream, +2 tool streams) to your live key and extends the subscription by 30 days. Same $49 price, no new key needed.