What is token compression?

Token compression is a technique that reduces the number of tokens sent to an LLM during inference by removing low-value tokens while preserving the information needed to answer the current query. It's not summarization — it selectively removes lines that score low against the current question.

How much does SuperCompress cost?

SuperCompress is free and open source under the MIT license. The Python library costs nothing. The hosted API has a free tier with 100K tokens per month and paid plans starting at $10/month.

What is LLM context compression?

LLM context compression reduces the size of the prompt sent to a language model by removing irrelevant or low-value tokens while preserving the information needed to answer the current question.

How does SuperCompress compare to truncation?

Truncation keeps only the head and tail of context, dropping the middle. SuperCompress compiles context around the current query, removes low-value blocks, and reports how much important context was kept.

Does SuperCompress require a GPU?

No. SuperCompress runs entirely on CPU with a tiny ~5K parameter policy. It compresses context before GPU inference with ~60ms latency.

Is SuperCompress free and open source?

Yes. SuperCompress is open source under the MIT license. The Python library and browser demo are free. The hosted API has a free tier with 100K tokens/month.

Cut Your LLM Token Costs by 65%.

Reduce input tokens before OpenAI, Claude, Gemini, or local model calls. SuperCompress compresses chat history, RAG chunks, support transcripts, tool traces, and app context before inference while preserving answer-critical evidence.

Start saving now Get API key

Prompt compression · RAG context · chat memory · CPU-only policy

Consumer AI apps send chat history, retrieved docs, tool traces, and user context on every request. That context drives LLM API cost, latency, and context-window pressure before the answer is generated. SuperCompress runs before inference, selecting the context most relevant to the current user request. Use it for chatbots, AI search, support agents, copilots, RAG, and any feature where context grows with users.

Control AI feature cost before the model call

Paste the same context a production AI feature would send: chat history, retrieval output, support logs, or tool results. SuperCompress shows what can be removed before the request reaches the model.

Cost model

Estimate token savings before rollout

Compare original tokens, compressed tokens, percent removed, and estimated inference impact before adding compression to production traffic.

Tokens saved = original − kept per compression
Works with OpenAI, Claude, Gemini, and local model calls

Quality

Prompt compression without blind truncation

Head/tail truncation can remove the line that answers the user. SuperCompress selects context against the current query and reports important-kept metadata.

82.5% average token savings on bundled long-context presets
Important-kept and risk metadata on every compiler result

Deployment

Drop-in preprocessing for production AI apps

Run compression locally with the Python package or call the hosted API before sending context to the LLM provider.

pip install · MIT license
Live hosted API + dashboard

Try a consumer app context

User request Prompt context

Enter in question field · ⌘/Ctrl+Enter in context

Runs locally with the trained SuperCompress compiler. The user request defines what matters, so duplicate history, irrelevant retrieval chunks, and tool noise can be removed before the LLM call.

Token & inference impact

Input tokens —
Tokens after SuperCompress —
Tokens removed —
Tokens saved —
Context size saved —
Answer quality retained —
Estimated prefill compute (before) — est. prefill
Estimated prefill compute (after) — est. prefill
Estimated compute saved —
Cooling estimate saved — datacenter cooling (est.)
Emissions estimate saved —

Compress prompts before they hit the model.

SuperCompress is built for production AI features with growing context: chatbots, AI search, support copilots, RAG pipelines, and agent workflows. On bundled long-context presets, compiler mode removes 82.5% of tokens on average while keeping query-critical evidence.

Oracle recall at 35% budget: SuperCompress 100%, H2O ~98%, FIFO and truncation ~25% — Baselines SuperCompress Fixed 35% budget · 8 seeds

Compiler mode token savings on long-context presets — Query-aware compiler mode Real preset contexts

Quick start

curl -d "context=…&query=…" \
  https://supercompress.dev/compress \
  -H "X-API-Key: sc_live_…"

Production API — metered in your dashboard. Requires a sc_live_… key. Python: pip install supercompress · API reference