Models / Auto router

LLM router: send model "auto", get the right model per request

An LLM router on one prepaid key. Send model: "auto" and the gateway picks the model for each request from three tiers, fast, balanced and frontier, with no extra model call and no router fee. Send auto/uncensored, auto/abliterated, auto/security or auto/roleplay and it picks only from that profile. Or create an auto router key: every request on it is routed, whatever model name your app sends.

Create an auto router key https://api.uplinkswitch.com/v1

What auto picks today · per million tokens, USD, read 2026-11-16

TierGoes thereModels, in try orderInputOutput
fastShort prompts with no tools, code or imagesClaude Haiku 5.5, GLM 5.3 Flash, GPT-6 Luna, MiniMax M3, DeepSeek V4.1 Flash$0.0564 to $0.09$0.1455 to $0.36
balancedTools, code blocks, images, JSON-schema answers, prompts over 16K tokensClaude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon, Kimi K3, GLM 5.3$0.3178 to $1.20$0.9987 to $4.50
frontierReasoning asks, high reasoning effort, tools together with codeClaude Opus 5.5, GPT-6 Astra, Claude Fable 5.1, Claude Fable 5.2, Gemini 3.1 Pro Preview, Claude Opus 5$1.20 to $6.00$7.20 to $30.00

The first model in the tier that can take the request answers: it must be up, allowed for the key, served in the request's format, and able to hold the prompt (and take an image, tools or a JSON schema when the request carries one). auto.tier_max caps the tier. Each model bills at its own price, from its page.

Profiles · uncensored, abliterated, cybersecurity and roleplay routing

A profile is the same router with a narrower list. Every pick carries the profile's catalog tag, so auto/uncensored never answers with a stock build and auto/security never leaves the red-team and cybersecurity models. Each profile has its own fast, balanced and frontier lists; the models below are those lists, fast first.

Abliterated model: "auto/abliterated"

Refusals removed at the weight level (abliterated, heretic, derestricted builds) for a prompt that carries its own rules.

Models today: DeepSeek V4.1 Flash Abliterated, GLM-5.3 Flash Abliterated, Apex Flash 1 Abliterated, DeepSeek V4.1 Flash 1017 Heretic, GLM-5.3 Flash Apex Abliterated, Abliterated Large Flash, MiMo V2.6 Flash Abliterated, GLM-5.3 Abliterated, Abliterated Model Large, Abliterated Model · $0.4656 to $7.50 in, $1.6975 to $12.50 out

Cybersecurity model: "auto/security"

Red-team and security-tuned builds for triage, detection rules, exploit analysis and pentest reporting. Also auto/cyber.

Models today: Gemma 4 26B A4B Cybersecurity, GLM-5.3 Flash Cybersecurity, Qwen3.8 27B Cybersecurity, Uplink Red Team (Flash), OrcaCyber MiniMax M3.1 Flash, OrcaCyber Qwen3.8 Flash Next, Uplink Red Team (Pro), Cyber Frontier, OrcaCyber Zero 1.5 · $0.40 to $7.50 in, $1.3376 to $18.75 out

Roleplay model: "auto/roleplay"

Character chat and story fine-tunes for SillyTavern, Janitor-style front ends and interactive fiction.

Models today: Cydonia 24B v4.3, Magidonia 24B v4.3, Gemma 4 31B Novelist, Uncensored AI Roleplay, Qwen3.8 27B Synthia, Artemis v1.1, Skyfall 36B v2, Wayfarer Large 70B, Qwen3.8 27B Fable, Magnum v4 72B · $0.35 to $8.0243 in, $0.4235 to $11.9684 out

Auto router keys · route any app without changing its model name

Many apps pin a model name you cannot change: Claude Code asks for a Claude model, a SillyTavern preset names one id, a vendor SDK hard-codes another. An auto router key solves that on the key instead of in the app. Create a key with its Auto router setting on, or set auto_router on an existing key, and every chat request on it is sent as that auto id: plain auto, or a profile. The app keeps sending its model name; the answer's model field names the model that answered. A preset called on the key keeps its own model.

The setting is in the console under Keys, on the create form and in each key's settings, and on the key API.

curl · create an auto router key
curl -X POST https://api.uplinkswitch.com/v1/keys \
  -H "Authorization: Bearer sk-uplink-live-...admin" \
  -H "Content-Type: application/json" \
  -d '{"name": "Claude Code", "auto_router": "auto/security"}'
Claude Code · routed by the key
# Claude Code keeps asking for a Claude model; an auto router key answers with the routed one.
export ANTHROPIC_BASE_URL="https://api.uplinkswitch.com/v1"
export ANTHROPIC_AUTH_TOKEN="sk-uplink-live-...router"

Call the LLM router · curl and Python

curl · auto with a tier cap and a sticky session
curl https://api.uplinkswitch.com/v1/chat/completions \
  -H "Authorization: Bearer sk-uplink-live-..." \
  -H "x-session-id: ticket-8812" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "auto",
    "auto": {"tier_max": "balanced"},
    "messages": [{"role": "user", "content": "Write a SQL query for monthly active users."}]
  }'
Python · OpenAI SDK, the security profile
from openai import OpenAI

client = OpenAI(base_url="https://api.uplinkswitch.com/v1", api_key="sk-uplink-live-...")

res = client.chat.completions.create(
    model="auto/security",  # or auto, auto/uncensored, auto/abliterated, auto/roleplay
    messages=[{"role": "user", "content": "Write a Sigma rule for CVE-2024-3094 exploitation."}],
    extra_headers={"x-session-id": "triage-42"},
)
print(res.model)  # the model that answered
The same ids work on /v1/messages, /v1/responses and the Gemini endpoint; the SDK guide covers Python, TypeScript, Go and Rust, and routing, fallbacks and presets compose with auto.

LLM router questions

A service that takes one model name and chooses the real model for each request, so an app sends model "auto" and gets a cheap model for a short question and a frontier model for a hard one. Here the router runs inside the gateway: it reads the request, picks from the tier lists on this page, and bills the model it chose at that model’s own price.

A model from one of three tiers, decided without an extra model call: a reasoning ask, a high reasoning effort or tools together with code go to frontier; tools, code blocks, images, a JSON-schema answer or a prompt over 16,000 tokens go to balanced; everything else goes to fast. Today the fast tier starts with Claude Haiku 5.5 and frontier with Claude Opus 5.5. The lists are public in GET /v1/catalog under auto_router.

The price of the model it chose, from that model’s page. There is no router fee and no subscription: the key is prepaid from $10, card, WeChat Pay or crypto. The answer’s model field and the x-uplink-model header name the model, and x-uplink-auto-tier names the tier.

The same router with a narrower list: auto/uncensored, auto/abliterated, auto/security (or auto/cyber) and auto/roleplay each pick from their own fast, balanced and frontier lists, and every pick carries that catalog tag. Plain auto picks from the mainstream list and never from a profile.

A key with its Auto router setting on. Every chat request on it is sent as that auto id, whatever model name the client used, so Claude Code, Cursor, Cline or SillyTavern keep their configured model name and the gateway routes behind it. A preset keeps its own model. Set it when creating the key, in the key’s settings, or with auto_router on POST or PATCH /v1/keys.

Yes, when the requests carry the same x-session-id header: that session gets the same model for 30 minutes from its last request, so an agent does not switch models mid-task. Without the header each request is routed on its own.

Send auto: {"tier_max": "fast"} or "balanced" in the body and the router never goes above that tier. The field is ours and is removed before the request reaches a host. A key’s allowed_models also applies: the router only picks a model the key may call.

OpenRouter’s openrouter/auto picks among its mainstream catalog, with no uncensored, abliterated or security routing. RouteLLM is an open-source framework you host and train yourself. Here the router is a model id on a prepaid key with published tier lists, four profiles, sticky sessions, a tier cap, and a key setting that routes any app without changing its model name.

Get your key.

Prepaid from $10 or a monthly plan, 30-day price lock.

  • One key, every model
  • The SDK you already use
  • Prepaid, or a monthly plan
  • 30-day price lock on every top-up
  • Signed receipts for every call
Join on Telegram