Does output efficiency injection reduce reasoning quality?

No. The governor targets polite fluff, redundant conversational preambles, and excessive formatting without constraining algorithmic depth, code accuracy, or logical reasoning steps.

What LLMs are supported by the dual-stage architecture?

All major frontier models including OpenAI GPT 6 Astra, GPT-5.6 Sol, Anthropic Claude Fable 5.1 and Opus 5, DeepSeek R1, Kimi K3, and Qwen 3.8 are fully supported.

Core Engine Architecture

Dual-stage AI token optimization engine

Simultaneously compress input context payloads and eliminate expensive output generation multipliers with zero codebase modifications.

Combined Net Savings

Up to 65%

Blends 50-68% input context reduction with 30-50% output generation brevity.

Proxy Latency

< 15ms

Ultra-fast in-memory streaming gateway executing at the edge.

Integration Effort

60 Seconds

Drop-in baseURL configuration with zero prompt refactoring required.

The production engineering challenge

Modern AI applications suffer from asymmetric billing: input tokens carry massive context payload bloat from RAG docs and system instructions, while output tokens cost 300% to 500% more per million. Optimizing only input or only output leaves over half of potential savings on the table.

Technical architecture & mechanics

Stage 1: In-memory input context compression

Analyzes prompt AST and natural language payloads to eliminate token-heavy sub-words, repetitive delimiters, and conversational fluff while preserving code blocks bit-for-bit.

Stage 2: Output efficiency governor

Dynamically injects behavioral brevity directives into API calls that direct the target model to omit polite conversational preamble, repetitive confirmations, and markdown fluff.

End-to-end transparent reverse proxy

Operates as a high-speed gateway requiring only a single baseURL change in your OpenAI, Anthropic, or Vercel AI SDK client configurations.

Standard OpenAI SDK Drop-in

import OpenAI from 'openai'; // Point standard OpenAI SDK to Curtly dual-stage proxy const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY, baseURL: 'https://proxy.curtly.dev/v1', defaultHeaders: { 'x-curtly-key': process.env.CURTLY_API_KEY, 'x-curtly-efficiency': 'balanced', // lite | balanced | ultra }, }); const response = await client.chat.completions.create({ model: 'gpt-6-astra', messages: [{ role: 'user', content: 'Analyze this customer dataset...' }], });

Architecture comparison

Evaluation metric	Curtly.dev gateway	Traditional / naive approach
Input Context Compression	Automated 50-68% reduction with AST Safe Vault lock	Manual prompt shortening or static truncation
Output Token Governance	Automated 30-50% output reduction eliminating 5x multipliers	None (full price paid on verbose output generation)
Syntax Preservation	100% bit-for-bit guarantee on code and JSON	High risk of JSON corruption and lost code delimiters
Integration Overhead	1-line baseURL swap in standard SDKs	Complex custom middleware and prompt rewrites

Frequently asked questions

Basic prompt compression only reduces the input prompt before sending it to the model. Dual-stage optimization combines deterministic input compression with active output efficiency injection, targeting the disproportionately expensive 3x to 5x output token rates.

Explore related platform features

Context Window Optimization

Ready to optimize your AI infrastructure?

Start optimizing in 60 seconds with a 1-line baseURL change. Free tier available with zero credit card required.

Start optimizing free Simulate token ROI