LIVE · TAO
TAO$— SUBNETS VALIDATORS256
Bittensor intelligence updates
Home / Interviews/ SOMA (SN114) Prepares for Early…
INTERVIEWS

SOMA (SN114) Prepares for Early Access Launch

SOMA (SN114) launches early access this month with a working proxy layer that cuts Copilot token bills by 10% to 30% using zero-code context compression

SOMA (SN114) Prepares for Early Access Launch

Every coding agent bill climbing month after month is a problem SOMA (SN114) has quietly been solving through bi-weekly competitions since January.

In an appearance on TAO PIll Podcast, the team announced that early access opens later this month, with SOMA (SN114) already delivering measurable Copilot savings.

Compression sits inside the agent to LLM pipeline, stripping bloated context down to what the model needs before each prompt gets sent.

10% of a monthly Copilot bill is today’s benchmark, while the team wants to return three months later and announce 30% instead.

How the Proxy Layer Works

SOMA (SN114) sits directly between a coding agent and the LLM it queries, with every prompt passing through the compressor before it reaches the model.

SOMA’s Network Dashboard

1. The agent sends its full context to SOMA (SN114): Chain of thought, prior conversation, code files, everything the agent thinks it needs.

2. SOMA (SN114) compresses the payload: Only the most crucial parts of the context get forwarded to the model.

3. The LLM responds normally: The model has no idea a compression step happened.

4. The response returns to the agent: The full loop continues, but the token bill for that turn shrinks noticeably.

5. Integration is a single bash command: No architectural changes on the customer side, no code refactoring, no infrastructure work.

The proxy approach is what unlocks closed-source coding agents like Copilot, Claude Code, and Codex. Those tools cannot be modified by outside developers, so the only way to compress their context is to intercept prompts on the way out.

The Track Record So Far

SOMA did not arrive at the Copilot proxy layer on day one. The team worked its way through progressively harder targets, benchmarking against the incumbent at each stage.

1. Plain text compression came first: SOMARIZER, the plain-text tool, is already live and beat LLMlingua (Microsoft’s state-of-the-art plain text compressor) on their internal QA benchmark.

SOMARIZER’s Web Interface

2. OpenFlow (open-source coding agent) came next: SOMA (SN114) open-sourced its OpenFlow plugin after miners pushed real-world savings to roughly 50% on token bills, deciding the problem was solved cleanly enough to release publicly.

3. Copilot is the current battleground: The subnet is on its third bi-weekly competition against Copilot, with miners producing roughly 15% savings on the internal benchmark and roughly 10% in genuine daily use.

4. Every evaluation runs twice: Once as a baseline without compression, once with the miner’s compressor active. Quality has to be preserved before token savings even get counted.

5.  Task runs happen 30-40 times each: Repeated evaluation removes the noise from LLM non-determinism and keeps rankings honest.

The miner incentive design carries this specific insight: a compressor that saves tokens but degrades output quality is worth zero to a paying customer, and gets scored accordingly.

The Revenue Model

SOMA is not planning to sell a fixed monthly subscription. The commercial model is designed to align directly with the savings it delivers.

1. Customers get an API key: Top it up like any credit-based service, plug it in with one bash command, use it with any provider they already work with.

2. SOMA (SN114) collects a percentage of the savings: If a customer saves $10,000 a month on token bills, SOMA takes a share of that number rather than a flat fee.

3. Roughly 40% of savings is the current working number: Not final, but the shape of the deal is clear.

4. 0% miner burn on the subnet: Everything flows to contributors right now while the product finds its market.

Combined with 100,000 SN114 alpha tokens locked in Conviction and no SN114 sold outside a single OTC deal with Const, the token side of the story is aligned with product execution rather than short-term selling pressure.

Signing Up Before the Rush

Early access opens at the end of this month, with invitations already live for anyone wanting to join the first wave of testers. The team is testing internally at Dendrite, the company behind SOMA, and expects user feedback to push compression rates significantly higher through fall.

Request for Early Access Into SOMA

Every coding agent user, from solo Copilot developers to teams spending five figure sums monthly across engineering, has reason to watch the release closely. 10% off a token bill is meaningful money, but 30% off turns the savings into a defensible product.

Enjoyed this article? Join our newsletter

Get the latest TAO & Bittensor news straight to your inbox.

We respect your privacy. Unsubscribe anytime.

The Daily Dispatch

Enjoyed this article?
Join our newsletter

Get the latest TAO & Bittensor news straight to your inbox — every morning before markets open.

IA
Ige A
Senior Editor

Be the first to comment

Leave a Reply

Your email address will not be published.


*