Reduce the tokens your models consume without compromising the quality of their output. Lower costs, faster responses, more usable context.
Context compression is the process of reducing the number of tokens required to produce a given model output, without compromising the quality of that output.
A long input goes in. A shorter, equivalent input comes out. The model receives less to process, costs less to run, and responds faster. The downstream task performs the same.
Hybrid approaches combining multiple compression techniques tend to produce the strongest results in practice.
Compression families
Extractive
Identifies and keeps the highest-signal tokens, removing the rest while preserving the original wording.
Abstractive
Rewrites the input as a shorter summary that conveys the same operational meaning.
Learned encoders
Transforms the input into compact representations the model can interpret directly.
Compression is one of the highest-leverage optimizations available to any team running large language models in production.
Input tokens dominate cost
Input tokens, not output tokens, drive the overwhelming majority of total spend in agentic workloads - even with caching enabled. Reducing input scales directly into savings.
Long contexts degrade output
Models pay less attention to information in the middle of long contexts. Accuracy often peaks at intermediate token counts and degrades as context grows. Compression is frequently sharper, not just cheaper.
Latency scales with input
Every additional thousand tokens extends model response time. For interactive applications and real-time agents, compression translates directly into a faster product.
Demand is accelerating
Programming alone grew from 11% to over 50% of all LLM token usage on OpenRouter between early and late 2025. Agentic workflows, RAG pipelines, and document analysis are driving sustained demand.
SOMARIZER delivers context compression as a continuously improving service. The architecture is designed to surface the strongest available implementation at any moment, automatically.
Static methods age out
A compression method tuned for one task family fails on another. The moment underlying models change, static approaches become liabilities.
Continuous improvement
SOMARIZER treats compression as an open and continuous problem. Users always receive the current best implementation with no engineering work to upgrade or swap providers.
Submit
The platform receives the input context along with the downstream task it needs to support.
Compete
Independent providers run their own compression implementations against the input - extractive, abstractive, learned encoders, or hybrids of all three.
Score and return
An evaluation layer measures every output against two criteria: compression ratio and fidelity to the original task result. The strongest output is returned. The strongest providers earn a larger share of the network.
SOMA is a token compression layer for AI agents and LLM pipelines. It reduces the number of tokens sent to the model - cutting inference costs without degrading output quality. SOMA specializes in two compression types: Context Compression, which strips redundancy from long inputs before they reach the model, and CoT Compression, which condenses an agent's accumulated session history so multi-step workflows stop resending tokens the model doesn't need.
Context Compression removes redundancy from long prompts, documents, and RAG inputs before they reach the model. The model sees a shorter, semantically equivalent input - which means fewer tokens billed and more usable space inside the context window. For high-volume AI workloads, the savings compound with every request.
Agents accumulate context as they work: every reasoning step, tool call, file read, and result gets appended to the session history - and resent to the model on every subsequent call. CoT Compression intercepts that session log before each call, scores each unit by relevance to the current task, removes redundancy and noise, and reconstructs a compact prompt that retains the state the model actually needs. The agent keeps its full working memory; the model stops paying for the parts that don't matter.
SOMA runs as a decentralized competitive network. Contributors submit compression implementations that are continuously benchmarked on both compression ratio and output fidelity. Only implementations that reduce tokens while maintaining quality earn rewards - so the network self-selects for compression that actually works.
SOMA runs as a proxy between your agent and its LLM provider. Integration is a single command: point your agent at the SOMA endpoint instead of calling the provider directly. SOMA compresses the context and forwards the request to the model of your choice - the response flows back unchanged. No infrastructure changes, no model retraining, no changes to your agent's logic. Get an API key on the SOMA Access page and you can be compressing on the next request.