🚀 OpenRouter Ultra V1.1

A production-ready AI gateway built for reliability, structured outputs, and cost-efficient LLM workflows.

OpenRouter Ultra is not just another proxy. It is an intelligent execution layer on top of OpenRouter that makes AI responses usable in real-world systems.

---

⚡ Why OpenRouter Ultra?

Most LLM APIs break in production because of:

• ❌ invalid JSON
• ❌ inconsistent outputs
• ❌ high token costs
• ❌ unpredictable latency
• ❌ no observability

Ultra solves this.

---

🧠 What makes Ultra different

✅ Reliable Structured Outputs

Ultra includes a multi-stage repair + validation pipeline:

• removes prose and markdown wrappers
• fixes quotes and trailing commas
• balances brackets
• enforces JSON schema

→ You get usable data instead of broken responses

---

⚡ Smart Routing (Speed, Cost, Quality)

Ultra automatically selects the best model based on:

• task type (chat, extraction, coding, etc.)
• latency targets
• cost constraints

Available presets:

• ultra-fast → lowest latency
• ultra-smart → best quality
• ultra-private → privacy-focused
• ultra-json → structured output optimized
• ultra-batch → high-throughput processing

---

💸 Cost & Budget Control

Set limits directly in your request:

• max cost per request
• max tokens
• model fallback strategies

→ prevents runaway LLM costs in production

---

🔍 Full Observability

Every request returns:

• latency
• selected model
• routing decisions
• repair status
• estimated cost

Plus debug endpoints:

• /v1/request/:id
• /v1/debug/:id

---

🧪 Built-in Repair Engine

Ultra automatically fixes common LLM issues:

• missing quotes in JSON
• invalid syntax
• trailing commas
• broken structures

→ dramatically reduces parsing failures

---

🔁 Optional TOON Format (Token Optimized)

Ultra supports an alternative format for structured data:

• less verbose than JSON
• fewer tokens for tabular data
• easier for LLMs to generate correctly

Content-Type: text/plain
inputFormat: "toon"
outputFormat: "toon"

Use when working with large structured payloads or batch jobs.

---

🧩 API Overview

Base Endpoint

POST /v1/ultra

Example Request

{
  "input": "Extract name and age from: John is 32 years old",
  "mode": "ultra-extract",
  "jsonSchema": {
    "type": "object",
    "properties": {
      "name": { "type": "string" },
      "age": { "type": "number" }
    }
  }
}

---

Example Response

{
  "success": true,
  "data": {
    "name": "John",
    "age": 32
  },
  "meta": {
    "latencyMs": 842,
    "model": "anthropic/claude-3-haiku",
    "costUsd": 0.0021,
    "repairApplied": true
  }
}

---

🔧 Specialized Endpoints

Endpoint	Description
/v1/ultra	Main intelligent routing endpoint
/v1/extract	Structured data extraction
/v1/convert	Format conversion (JSON ↔ TOON)
/v1/repair	Fix broken LLM outputs
/v1/request/:id	Retrieve past request
/v1/debug/:id	Debug full pipeline

---

🧠 Core Concepts

1. Ultra Modes

Define how the system behaves:

• speed vs quality vs cost
• structured vs freeform
• batch vs realtime

---

2. Repair Pipeline

Every response passes through:

1. cleanup
2. syntax fixing
3. structure repair
4. schema validation

---

3. Task Awareness

Ultra detects intent:

• extraction
• summarization
• chat
• coding

→ routes to optimal model chain

---

📦 Use Cases

• AI-powered SaaS backends
• data extraction pipelines
• automation workflows
• marketplaces & platforms
• batch processing jobs
• agent systems

---

🚀 Why this matters

Most developers spend hours fixing:

• broken JSON
• inconsistent outputs
• model switching logic

Ultra handles this automatically.

→ You focus on your product → Ultra handles the AI complexity

---

🔑 Authentication

Use your OpenRouter API key:

$Authorization: Bearer YOUR_OPENROUTER_KEY

---

⚙️ Running on Apify

This actor runs as a serverless AI gateway:

• scalable
• production-ready
• easy to integrate
• supports batch + realtime

---

🤝 Contributing / Feedback

If you're building serious AI systems, feedback is welcome.

---

⚡ TL;DR

OpenRouter Ultra turns:

LLM output → broken JSON → debugging hell

into:

LLM output → validated structured data → production-ready