Prompt compression for OpenRouter applications

Keep OpenRouter. Send it smaller prompts.

Compress conversation history, RAG results, tool output and application context before OpenRouter routes the request—reducing input tokens and helping the model focus on what matters.

Test an OpenRouter prompt View the integration

Start free · No saving, no compression charge · See pricing

Before1,248 tokens

After782 tokens

Required content preserved466 saved tokens

Lower input costs

Send fewer billable input tokens through OpenRouter or directly to your model provider.

Faster long-context responses

Reduce the amount of context the model has to process on suitable workloads.

Better model focus

Remove duplicated and low-value context so important instructions and facts stand out.

Before and after

See what the model no longer needs to read.

Original

1,248

tokens

Compressed

782

tokens

Removed

37%

of input

Repeated background and low-value restatements are removed or rewritten.

Protected instructions, values and identifiers remain unchanged.

Results vary by prompt, workload and model.

The outcome is visible without playback; the short product film shows the transformation in action.

Continue in a private compression workbench.

Run up to 100 volatile-prompt compression tests plus progressive Prompt Review for stable instructions without creating an account.

Open the free trial

Production lesson · FocusedFit.ai

The best compression result was fixing the application.

FocusedFit.ai tested Runtime Compression in shadow mode against a roughly 150 KB context payload. Inspection exposed truncated source data and duplicated raw and structured content.

After fixing the application, the payload fell to roughly 40 KB by size. UsageTap estimated only another 3–5% reduction from Runtime Compression, so the team left it off.

Read the case study →

Context investigated: ~150 KB
Tightly structured payload: ~40 KB
Further reduction estimated: 3–5%

Good optimization isn't about maximizing compression. It's about sending the model only what the task needs.

Application context first

Built for the context AI applications actually send.

Conversation history

Reduce growing histories in assistants, support products and copilots.

Agent and tool output

Trim verbose tool responses, logs and repeated execution context before the next model call.

RAG results

Remove repetition and low-value retrieved text while retaining important facts, instructions and citations.

Application context

Compact structured records, tenant configuration, JSON and other application state included in prompts.

Also useful forLong documentsTranscriptsOther text-heavy inputs

Provider-neutral integration

Keep your provider. Send it smaller prompts.

Choose OpenAI, Anthropic, UsageTap Gateway, or OpenRouter. Add the UsageTap SDK, call Compress over HTTPS, or use the same language-agnostic flow from Python.

View SDK setup options →

Provider or route

Integration style

Your selection

Keep the OpenRouter SDK and add the UsageTap SDK to compress context.

Install

npm install @usagetap/sdk openai

OpenRouter · UsageTap SDK

import OpenAI from "openai";
import { withCompression } from "@usagetap/sdk/openrouter";

const provider = new OpenAI({
  baseURL: "https://openrouter.ai/api/v1",
  apiKey: process.env.OPENROUTER_API_KEY,
});

const openrouter = withCompression(provider, {
  apiKey: process.env.USAGETAP_API_KEY,
  minContextTokens: 1_000,
});

const response = await openrouter.chat.completions.create({
  model: "anthropic/claude-sonnet-4",
  messages,
});

Open Code Integration SDK reference

Simple prepaid pricing

Start free. Pay only for tokens Compress removes.

No subscription or credit card is required. Real-time Compress is part of the same prepaid UsageTap Balance as Meter and optional AI-powered work.

See all UsageTap pricing →

Free allowance: 5M tokens saved/month
After the allowance: $0.30 per 1M tokens saved
No saved tokens: No Compression usage charge

A category-level comparison

Not all prompt reduction works the same way.

Category	Context-window truncation	General text compression	UsageTap Compress
Primary purpose	Make an oversized request fit within a model limit	Shorten prose such as documents, websites and transcripts	Optimize dynamic context assembled by AI applications
Common inputs	Messages selected by position	Natural-language text	Conversation history, RAG, tool output and structured application context
Protected content	Content may be removed based on location	Depends on the implementation	Explicitly protect important instructions, values, identifiers and structures
Skip decision	Primarily triggered by context limits	Often attempts compression on the supplied text	Can skip unsuitable or uneconomic candidates
Inspection	Limited visibility into removed context	Depends on the implementation	Inspect original input, output, token counts and transformations
Fallback	Removed content is not sent	Depends on the implementation	Use the original input when validation fails

Runtime safeguards

Protect, inspect, and fall back.

Protect exact instructions, identifiers, values, and structures; inspect the result; and use the original input when validation fails.

Protect what matters

Keep exact instructions, identifiers, numbers, code and structured content intact.

Skip poor candidates

Do not compress prompts that are already efficient or unsuitable.

Inspect every result

See the original input, compressed output, token counts and transformations.

Use the original when checks fail

When validation fails, UsageTap sends the original prompt.

Open the LLM cost-reduction recipe book →Explore the SDK reference →

Workload evaluation

Prove it on your own prompts.

We’ll evaluate 50–100 representative OpenRouter requests and show saved tokens, compression latency, end-to-end latency, quality results, and fallback rate before you change production traffic.

Get an OpenRouter prompt review

Savings calculator

Model your prompt savings.

Your first 5M saved tokens are free each month. After that, UsageTap deducts $0.30 per 1M saved tokens. Compare that charge with your downstream input-token cost to estimate the net saving for your workload.

Downstream model

Loading the latest model prices…

Monthly input tokens25M

1M1B

Reduction assumption

Choose an assumption for your workload. Results vary.

Your input

25M

tokens

Saved

6.25M

tokens

Sent after compression

18.75M

tokens

Billable saved

1.3M

after free allowance

Original downstream input cost: —
Input cost after compression: —
Estimated input-cost savings: —
UsageTap Balance deduction$0.30 per 1M billable tokens saved: —
UsageTap Compress costPublished saved-token price is all-in: —
Estimated net savings after UsageTap Compress: —

Included Compress access is limited to 30 requests per minute. Custom terms can include different operating limits.

Practical questions about Compress

How is UsageTap different from OpenRouter context truncation?+

Context truncation helps an oversized request fit by removing messages or content, commonly based on position or limits. UsageTap is an additional optimization layer that evaluates eligible context before routing, protects selected content, exposes the transformation and can skip poor candidates.

Does UsageTap replace OpenRouter?+

No. Keep your OpenRouter client, model choices, routing and provider fallbacks. UsageTap wraps the client and sends OpenRouter a smaller prompt when compression is suitable.

Does compression always make a request faster?+

No. Sending fewer tokens can reduce model processing time, especially for long-context workloads, but compression itself also takes time. The net result depends on the prompt, model and deployment, so evaluate end-to-end latency on representative requests.

Can prompt compression improve response quality?+

Removing repetition and irrelevant context can help a model focus on important information, but it is not a universal quality improvement. Quality must be evaluated against the specific application workload.

What content can UsageTap protect?+

You can protect exact instructions, identifiers, numbers, code, structured data, citations and other spans your application needs to preserve.

What happens when a prompt should not be compressed?+

UsageTap can skip prompts that are already efficient, unsuitable or uneconomic. If validation fails, your integration can use the original input instead.

Can I inspect the compressed prompt?+

Yes. Inspect the original input, compressed output, token counts, decision metadata and transformations before deciding how to use the result.

How is Compress priced?+

The first 5 million saved tokens each month are free. After that, Compress deducts $0.30 of UsageTap Balance per 1 million saved tokens. The published saved-token rate is all-in; no additional UsageTap percentage fee is added. Skipped requests, unchanged results, compression errors and fallback to the original prompt record zero saved-token usage.

Is UsageTap Compress separate from UsageTap Meter?+

Yes. Compress works on its own. Use Meter for customer usage economics, Compress for prompt input efficiency, or both when your application needs both capabilities.

Make your next LLM request smaller.

Open the free compression trial Use with OpenRouter

Start free · No saving, no compression charge · See pricing