How does relayswitch/auto choose a model?

It classifies each request locally, with no extra model call. A reasoning ask, a high reasoning effort, or tools with code goes to frontier; tools, code blocks, images, a json_schema answer or a prompt over 16,000 tokens to balanced; the rest to fast. auto.tier_max caps the tier. The answer names the model and x-uplink-auto-tier its tier.

Will an agent switch models mid-run?

Not with a session id. Send x-session-id and the same model answers that session for 30 minutes from its last request.

When does a fallback model answer?

Only after every route of the model before it has failed before a byte reached you, so a stream falls back only before its first byte. Up to 5 models, tried in order. The one that answers is named in model and x-uplink-model and billed at its own price; the key’s allowed-model list applies to each.

A model, fallbacks, default params, a system prompt (prepended to yours or replacing it) and tools, under one name. Call it as model: "@preset/ ". Your request’s own fields win. Up to 100 presets per organization.

What do throughput and latency routing change?

The order a model’s routes are tried in: price (the default), latency (quickest first token lately) or throughput (most output tokens per second lately). The price you pay is the same whichever route answers. OpenRouter-style provider.sort is read the same way.

Can I limit how much an agent runs at once?

Yes, per key: concurrency (requests in flight, 1 to 10,000), a spend limit that resets daily, weekly or monthly, Requests / min and tokens / min.

Routing

The right model per request, and a backup when it fails.

Send relayswitch/auto and the router picks a fast, balanced or frontier model for each call. Name fallbacks, save presets, route by throughput. Every answer says which model it was and bills at that model's price.

Get an agent key Code

Auto router, live

Model id: relayswitch/auto
Fast tier: 5 models
Balanced tier: 5 models
Frontier tier: 5 models
Fallbacks per request: up to 5
Session stickiness: 30 minutes

What it does for an agent

Auto router: relayswitch/auto classifies the request on the API and picks a tier. No extra model call, no added fee.
Sticky per session: With x-session-id the agent keeps one model for the whole run, so its context and tool habits stay put.
Fallback lists: models: [...] names up to 5 backups. A key can carry fallback_models for every request that names none.
Saved presets: Model, fallbacks, params, system prompt and tools under one name, called as @preset/<name>. Change the preset, not the agent.
Throughput routing: routing: throughput tries the route with the most tokens per second first; latency, the quickest first token. Same price.
Spend and concurrency per key: Daily, weekly or monthly spend limits and a cap on requests in flight, so a runaway agent stops at its own line.

Copy, paste, run

OpenAI SDK: auto router

python

1from openai import OpenAI

2import os

4client = OpenAI(

5 base_url="https://api.relayswitch.io/v1",

6 api_key=os.environ["RELAYSWITCH_API_KEY"],

7)

9reply = client.chat.completions.create(

10 model="relayswitch/auto",

11 messages=[{"role": "user", "content": "Refactor this function to async."}],

12 extra_body={"auto": {"tier_max": "balanced"}}, # never frontier

13 extra_headers={"x-session-id": "agent-run-42"}, # same model all run

14)

15print(reply.model) # the model the router chose; billed at its price

Anthropic SDK: fallbacks and throughput

python

1import anthropic

2import os

4client = anthropic.Anthropic(

5 base_url="https://api.relayswitch.io/v1",

6 api_key=os.environ["RELAYSWITCH_API_KEY"],

7)

9reply = client.messages.create(

10 model="claude-sonnet-5-5",

11 max_tokens=1024,

12 messages=[{"role": "user", "content": "Plan the migration."}],

13 extra_body={

14 "models": ["gpt-6.1-sol", "kimi-k3"], # tried in order if claude-sonnet-5-5 fails

15 "routing": "throughput", # fastest output first

16 },

17)

18print(reply.model)

curl: a preset

bash

1# Save the agent's model, fallbacks, defaults and system prompt once (admin key).

2curl https://api.relayswitch.io/v1/presets \

3 -H "Authorization: Bearer $ADMIN_KEY" \

4 -H "Content-Type: application/json" \

5 -d '{

6 "name": "coder",

7 "model": "claude-sonnet-5-5",

8 "models": ["gpt-6.1-sol"],

9 "params": {"temperature": 0.2, "max_tokens": 4096},

10 "system_prompt": "You are a careful senior engineer.",

11 "description": "Coding agent defaults"

12 }'

14# Then any agent calls it by name.

15curl https://api.relayswitch.io/v1/chat/completions \

16 -H "Authorization: Bearer $RELAYSWITCH_API_KEY" -H "Content-Type: application/json" \

17 -d '{"model": "@preset/coder", "messages": [{"role": "user", "content": "Fix the failing test."}]}'

Key API: limits and default fallbacks

bash

1# An agent key with a daily spend limit, 8 requests in flight and fallbacks.

2curl https://api.relayswitch.io/v1/keys \

3 -H "Authorization: Bearer $ADMIN_KEY" -H "Content-Type: application/json" \

4 -d '{"name": "nightly-agent", "monthly_cap_usd": 5, "spend_limit_period": "daily",

5 "concurrency": 8, "routing": "throughput",

6 "fallback_models": ["gpt-6.1-sol", "kimi-k3"]}'

The auto router’s tiers today

From GET /v1/catalog auto_router, priced live in USD per million tokens.

Tier	Model	Input	Output
fast	Claude Haiku 5.5	$0.06	$0.15
fast	GLM 5.3 Flash	$0.09	$0.15
fast	GPT-6 Luna	$0.0625	$0.20
fast	MiniMax M3	$0.09	$0.36
fast	DeepSeek V4.1 Flash	$0.06	$0.18
balanced	Claude Sonnet 5.5	$1.20	$3.00
balanced	GPT-6.1 Sol	$0.60	$3.00
balanced	Gemini 4 Argon	$0.60	$3.00
balanced	Kimi K3	$0.90	$4.50
balanced	GLM 5.3	$0.3276	$1.0296
frontier	Claude Opus 5.5	$2.80	$10.00
frontier	GPT-6 Astra	$3.00	$15.00
frontier	Claude Fable 5.1	$6.00	$30.00
frontier	Gemini 3.1 Pro Preview	$1.20	$7.20
frontier	Claude Opus 5	$1.80	$7.50

Agent setup guides

More features

FAQs

Questions

For agent builders.

Open the console

A key in under a minute

Get a key

Or email hello@relayswitch.io

Route your agents by the request

One model id that picks the tier, a backup list, and the price of the model that answered.

All features Get an agent key