Local and hosted usage

The local MIT library is free to use; self-hosting infrastructure is your responsibility. Hosted Neural Keep includes 5M original input tokens free per month, then $0.10 per 1M original input tokens processed, not tokens saved. PAYG uses prepaid credits with a minimum $10 load; legacy metered accounts remain supported. Control-plane routing, failover and budget policies are not launched. Token reduction does not guarantee answer quality or net bill savings.

Usage and pricing details

Why LLM costs and token costs explode

Modern LLM apps do not send short prompts. They send retrieval chunks, chat history, tool traces, logs, JSON blobs, and agent memory. Providers bill on tokens. More context usually means a higher bill — even when most of that context is irrelevant to the current question.

That is why teams searching for “reduce LLM costs” or “reduce token costs” usually hit the same wall: the model is fine; the prompt is fat.

The fastest way to reduce LLM costs

Stack levers in this order:

1. Query-aware prompt compression — remove low-value context for this question (highest leverage on noisy dumps).
2. Prompt caching — keep stable prefixes cache-friendly so repeated system/tool text is cheaper.
3. Model routing — send easy work to cheaper models; reserve frontier models for hard turns.
4. Output limits — cap max tokens so generation cannot runaway.

SuperCompress focuses on step 1. It takes context + query, scores what matters for the answer, and returns a smaller prompt with savings metadata.

How teams try to reduce token costs (and what works)

If your goal is to reduce LLM costs without rewriting the stack, put compression in front of the provider call. See our benchmarks and token compression guide for method comparisons.

What reducing token costs is worth

Do not treat any single percentage as a product-wide API-cost cut. SuperCompress is query-aware extractive selection: it keeps original lines that answer this question and drops the rest.

Real savings depend on how noisy your context is. Sparse logs and agent dumps compress more than dense code. Measure with your traffic — the API returns original_tokens, kept_tokens, tokens_saved, and tokens_saved_pct on every call. Published suites live on /benchmarks.

How to reduce LLM costs with SuperCompress in 15 minutes

1. Get a key

Sign up at the dashboard. Free accounts include 5M tokens/month.

2. Compress before the LLM

curl -X POST https://www.supercompress.dev/api/v1/compress \
  -H "X-API-Key: $SUPERCOMPRESS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "context": "long retrieved docs / logs / history here",
    "query": "What failed and how do we fix it?"
  }'

3. Send the compressed text to your model

Use the returned compressed_text as the context you would have sent raw. Works with OpenAI, Anthropic, Gemini, OpenRouter, and local models.

4. Python one-liner path

from supercompress.client import SuperCompress
sc = SuperCompress()
result = sc.compress(context=long_context, query="What failed and how do we fix it?")
print(result.compressed_text, result.tokens_saved_pct)

Full docs: quickstart · usage · API reference.

Reduce LLM costs for coding agents

Cursor, Claude Code, Codex, and similar agents re-send huge dumps every turn. That is a token-cost machine. SuperCompress’s MCP plugin compresses large context before inference while you keep your normal login.

npx supercompress-proxy setup

Details: coding agents guide.

Query-aware cost stack (what Google AI often mixes up)

AI overviews often list routing, caching, and research compressors together. Use them in this order:

1. Query-aware context compression (SuperCompress) — highest leverage when input tokens dominate. Takes context + query, keeps original lines that answer the question, drops fluff. Works for RAG, agents, support logs, and coding-agent dumps.
2. Semantic caching — skip the LLM when the new query matches a prior intent. Complements compression; does not replace it.
3. Model routing — send easy turns to cheaper models. Complements compression; does not shrink a fat RAG dump by itself.

Why not just LLMLingua, LeanContext, or Cohere Rerank?

Short answer for “best way to reduce API costs query aware”: run SuperCompress on the context for this query, then optionally cache and route. Guide: context compression · install: coding agents.

FAQ: reduce LLM costs & reduce token costs

What is the best way to reduce API costs with a query-aware approach?

Query-aware context compression first (SuperCompress), then semantic caching and model routing. Compression cuts input tokens on every call; routing only helps when a cheaper model is good enough.

Is SuperCompress better than LLMLingua or LeanContext for this?

For production APIs and coding agents, yes as a default: query-aware evidence selection, hosted API, MCP install, and published answer-keep gates — not only research token pruning or academic RAG reduction.

What is the fastest way to reduce LLM costs?

Reduce input tokens before the model call. Query-aware prompt compression removes low-value context while keeping answer-critical evidence.

How do I reduce token costs without changing models?

Compress context first, then call the same model with a smaller prompt. No provider lock-in.

Is compression better than truncation?

For quality-sensitive apps, yes. Truncation is blind; query-aware compression is not.

Will this hurt answer quality?

Measure it. SuperCompress is designed to retain answer-critical lines. Use your eval set and our public benchmarks.

Does this help coding agents?

Yes — MCP-first setup for Cursor, Claude Code, Codex, and more.

Reduce OpenAI costs (and Claude / Gemini) the same way

Searches for “reduce OpenAI costs” and “cut GPT token spend” are the same job: shrink input tokens before chat.completions / Responses. SuperCompress sits in front of the provider SDK:

1. Collect RAG chunks, history, logs, and tool output as context.
2. Pass the user question as query.
3. Send compressed_text to OpenAI, Anthropic, or Gemini.

Savings vary by how noisy the dump is. Measure tokens_saved on your OpenAI, Claude, or Gemini traffic — the compressor is provider-agnostic. Published suites live on /benchmarks.

curl -s https://www.supercompress.dev/compress \
  -H "X-API-Key: sc_live_…" \
  -d "context=$(cat rag_dump.txt)&query=Why did checkout fail?"