Can I disable output efficiency for specific requests?

Yes. Simply pass x-curtly-efficiency: off in your request headers to bypass output injection while retaining input context compression.

Generation Token Governor

Output generation efficiency and brevity governor

Eliminate conversational preamble, apologies, and formatting fluff from LLM responses to slash expensive generation token invoices.

Output Token Savings

30% - 50%

Eliminates conversational fluff across all responses.

Invoice Impact

Up to 5x ROI

Targeted directly at the most expensive tier of LLM rate cards.

Response Latency

Faster TTFT

Fewer generation tokens directly accelerate end-user streaming.

The production engineering challenge

Frontier model output tokens cost 300% to 500% more per million than input tokens. LLMs by default produce conversational padding ("Sure, I would be happy to help with that!"), repeated questions, and verbose sign-offs that bloat bills without adding technical value.

Technical architecture & mechanics

Zero-overhead directive injection

Injects concise behavioral constraints into prompt framing during proxy dispatch, commanding the LLM to deliver direct, unpadded answers.

Multi-tier efficiency controls

Offers selectable brevity levels (Lite -20%, Balanced -35%, Ultra -50%) tailored to conversational assistants, coding agents, or automated pipelines.

Format integrity preservation

Guarantees that requested structured formats (JSON, YAML, Markdown tables) maintain clean syntax without truncation.

Python OpenAI Client with Efficiency Header

from openai import OpenAI client = OpenAI( api_key="your-openai-key", base_url="https://proxy.curtly.dev/v1", default_headers={ "x-curtly-key": "your-curtly-key", "x-curtly-efficiency": "ultra" # lite | balanced | ultra } ) response = client.chat.completions.create( model="gpt-6-astra", messages=[{"role": "user", "content": "Explain binary search implementation."}] )

Architecture comparison

Evaluation metric	Curtly.dev gateway	Traditional / naive approach
Fluff Elimination	Automated injection calibrated per model family	Manual system prompt tweaking per project
Output Rate Multiplier Defense	Active defense on 3x-6x generation pricing	Unchecked spending on redundant output tokens
Response Speed	Faster completion due to fewer tokens generated	Slow token streaming due to unnecessary pleasantries

Frequently asked questions

No. The directives explicitly command the model to provide complete, working code implementations while stripping only the conversational commentary surrounding the code.

Explore related platform features

Core Engine Architecture

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