Metagraph CapitalInvestment brief · Technical edition · Confidential11 October 2026

SOMA (SN114)

SOMA sits between a coding agent and its model and trims the conversation the agent re-sends on every call. Its best-known claim is that this cuts an agent's token bill by about 15%; its own site labels that figure an estimate and says the measured saving on GitHub Copilot traffic is "far lower", about 10% in one run (thesoma.ai/compression). The product charges the model's cost with no markup, has no disclosed customer or revenue, and no customer money reaches the token. The token is a claim on a $17.7M-a-year emission stream and on whatever Dendrite, the Cyprus company that owns the subnet, later decides to charge.

Market cap$13.4M1,989,868 α at $6.72, block 9,260,669, 11 Oct V

Alpha issued per year$17.7M7,200 α/day at spot; 132% of market cap V

Customer revenueNone foundnone claimed; taorevenue inflow $0, 30 days V

Owner's share of alpha23%457,613 α = $3.1M; never sold from that key V

In short

The product works and is priced at cost. The SOMA App trims what a coding agent re-sends and bills DeepSeek's list price with no markup; the team's own measured saving is about 10%.
The alternatives cost little and ship by default. Provider caching bills repeated input at 2–3% of list, and every major harness and API now compacts old turns itself; SOMA's own page says that where caching hits, "caching is simply the better instrument".
No revenue and no customer, found or claimed. taorevenue shows $0 of owner inflow and $0 of burns over 30 days; the site, X and TAO Daily name no customer.
The owner holds 23% of alpha and has never sold from that key. Dendrite is a 50-engineer Bittensor company that also builds SN3 and SN97; eight people have ever committed to SOMA's code and three did in October.
Verdict: the token is a claim on $17.7M a year of issuance and on Dendrite's future pricing, not on a business. It is a bet that Dendrite will price the product and route some of that revenue to alpha; neither has begun.

SOMA's best-known claim, a 15% saving, is an estimate; the team's own measurement is about 10%

The claims that circulate about SOMA, who made them, and what the primary source showsChecked 11 Oct 2026 unless dated.
Claim	Who said it	What the check shows
SOMA cuts an agent's token cost about 15%	thesoma.ai benchmark table S	Every row is labelled "est.". The team's compression page: the measured figure on Copilot CLI traffic is "far lower"; one run showed about 10% and "was not a clean baseline". The home page adds: "Real figures land with the benchmark release." V
"Around 10% token savings" at the Copilot launch	@SomaSubnet, 28 Aug S	Consistent with the compression page. No per-customer or aggregate measurement is published. V
One prompt cut from 12,937 to 7,801 tokens, −39.7% input cost	TAO Daily card, 7–13 Sep 1	One team-supplied prompt, not reproduced. Where a provider cache already serves most of the prompt, a token count overstates the cost saving; SOMA's own page says so.
"Our compressor comes from an open subnet where anyone can beat the current one"	thesoma.ai S	The contest is open and paid (below). No published link ties a winning submission to the production proxy; the only released compressor, Dendrite's OpenClaw plugin, credits no miner. V
Integrated with Chutes, Say GM and Engy as inference providers; proxying traffic to Albedo	@SomaSubnet, 9 Oct S	Integrations, not paying contracts. Albedo (SN97) is another Dendrite subnet. Volumes not disclosed.

The product is a no-markup proxy for coding agents on three DeepSeek models, with no named customer

What SOMA sells, at what price, and to whomthesoma.ai and app.thesoma.ai/v1/models, read 11 Oct 2026. V
Item	Detail
SOMA App (open beta)	A proxy: a developer points GitHub Copilot CLI or Codex at SOMA instead of the model; SOMA trims the conversation history, tool output and reasoning, leaves the instructions and the user's first message untouched, and passes the call on. Claude Code "soon", Zed "planned"
Models	Three DeepSeek models: V4 Pro, V4 Flash and V4.1 Flash. No OpenAI or Anthropic model yet
Price	"Zero margin. You pay what the model costs us." DeepSeek V4 Pro input $0.92 per million tokens at list, $0.78 "with SOMA", est. −15%. $5 of credits for new accounts; card or TAO
How Dendrite earns	"We earn when compression earns its keep, not on the spread." No fee, share-of-savings term or enterprise price is published
Other products	Somarizer, a free browser summariser; an open-source OpenClaw compressor (MIT); "SOMA for Enterprise", private deployment, "in design"
Customers	None named. No usage, token-volume or revenue figure on the site, docs, dashboard or X; none claimed in TAO Daily's six SOMA articles

A 50-engineer company owns SOMA; eight people have ever committed to its code

Who builds SOMA, and how the team size checks against the codeLinkedIn and Discord read 10 Oct; GitHub API 11 Oct 2026.
Item	Detail
Company	Dendrite Quantum LTD, Cyprus (HE 475409), owner of SN114; also builds Teutonic (SN3) and Albedo (SN97); sister company BlockWise in Warsaw V
Founder	Piotr Barbachowski, Limassol; describes SOMA as "our product" and Dendrite as 50 engineers grown from five people who first mined TAO in September 2022 S
Named staff	Aleksander Muszyński (integrations, per the founder's 8 Oct post); Jan Różycki and Hubert Korzeniewski (miner workshop) S. On Discord, "Matt \| SOMA" (GitHub bahamajohn, the top committer) posts scoring changes; "oli \| SOMA" holds the subnet-owner role V
Team size, checked	Eight people have ever committed to SOMA's five public repositories; three committed in October. The 50-engineer figure is company-wide across three subnets and BlockWise V

Code changes to SOMA's main repository fell from 117 in July to 19 in September

Commits per month to github.com/DendriteHQ/SOMA, GitHub API, read 11 Oct 2026; October covers 1 to 11 October. The benchmark and shared-library repositories were also updated on 6 October. V

Published methods match or beat SOMA's measured saving, and the model providers now compact context themselves

Context compression for coding agents: token saving and effect on solved tasks, SOMA against published alternativesPapers and docs linked; read 2–6 Oct 2026. SOMA figures from its own site, 11 Oct.
Method	Who	Saving	Effect on tasks solved	Cost to use
SOMA App	Dendrite S	est. −15%; measured "about 10%" on Copilot	Not published	Model cost, no markup
Observation masking (drop old tool output)	JetBrains, 2025	cost about −50%	−2.0 to +2.8 points	Free, about 50 lines of code
AgentDiet	PKU/ByteDance, 2026	cost −21 to −36%	+1.5 to +2.0 points	Free
SWE-Pruner	2026	−23 to −38%	+1.2 to +1.4 points	Free, MIT
Anthropic compaction and context editing	Anthropic docs	Server-side	n/a	No separate fee; the summary call is billed as tokens
OpenAI server-side compaction	OpenAI docs	Server-side	n/a	Billing not stated in the docs

The methods that keep task success intact are simple: drop old tool output and prune file reads. SOMA's contest may find better ones; on its own hidden tasks the 3 October leader cut weighted tokens about 64%, but the product has not shown that gain on real traffic.

The alternatives to SOMA are included in what buyers already pay for

What a buyer can use instead of SOMA, and what each costsPricing pages read 11 Oct 2026 where linked; others 2 Oct.
Alternative	What the buyer gets	Price
Provider prompt caching	Repeated input billed at a fraction of list: Anthropic cache hits $0.25 against $10 per million input tokens on its top model (docs); DeepSeek cache hits $0.003 against $0.15 (pricing) V	Included; cache writes cost 1.25–2× list at Anthropic
Harness auto-compaction	Claude Code, Codex, Gemini CLI, Cursor and OpenClaw summarise or prune old turns themselves	Included in the harness; the summary is billed as tokens
Gateways	OpenRouter routes at provider price with sticky routing for cache hits (docs); LiteLLM, Portkey and Kong bundle open-source compressors	No token markup; enterprise tiers
Compression start-ups	Compresr (query-specific compression API), The Token Company (YC, per token removed)	Usage-priced; none publishes a tier above $0.30 per million tokens (2 Oct)

The market is large; the slice a paid compressor can charge for is thin, and winning customers is the constraint

The market SOMA sells into, the slice it can reach, and what comparable companies are worthMenlo Ventures (Dec 2025), company announcements; read 2 Oct 2026 for the thesis-check brief, labelled single-source where so.
Measure	Figure
Floor: what a compression vendor can charge today	No standalone compression vendor discloses revenue; public price points run $0.10–0.30 per million tokens removed
Ceiling: enterprise spend on model APIs	$12.5B in 2025, coding $4.0B of it (Menlo)
The slice SOMA can reach	The share of an agent's bill that is not already a cache read; on a cache-heavy harness that is roughly 5–35% of input cost, so a 30% cut in new content is about 10% of the bill (house arithmetic)
Companies doing this in-house	Cursor $29.3B, Cognition $10.2B, Factory $5B, OpenRouter acquired by Stripe for over $7B 1; for all of them compression is a feature, not a product

Who earns what: every dollar paid on SOMA is new issuance

Who is paid on SOMA, for what, and whether any of it is customer moneyChain, block 9,260,669 (11 Oct 2026), alpha $6.72 = 0.02324 τ × Kraken TAO $289.30; 7,200 α issued a day, protocol split 18/41/41. V
Who	What they do	How they are paid	Per day	Customer money?
Miners (three paid slots)	Submit one Python file that rewrites the message list a coding agent sends; it is run on SOMA's hidden coding tasks and scored on tokens saved while still solving the tasks	New alpha; the top three slots hold 75%, 15% and 10% of incentive today	$19.8k	No
Validators and stakers (nine validators)	Set the weights SOMA's platform returns after scoring. The owner's own validator takes 55% of dividends	New alpha	$19.8k	No
Dendrite, the subnet owner (owner coldkey 5FHrQ…)	Runs the contest platform, the hidden task set and the SOMA App proxy	The 18% owner cut in alpha	$8.7k	No
Alpha holders	No buyback, burn or fee sink. taorevenue, 30 days to 11 Oct: owner inflow $0, burn $0, outflow $1,138,744	—	—	No

The miner contest runs every two weeks and pays three slots

How SOMA's miner contest worksthesoma.ai/dashboard and chain, 11 Oct 2026; scoring from the repository docs, read 2 Oct. V
Question	Answer
Cadence	Two weeks per round: one week of uploads, one of evaluation. Round 11 ("CoT-Compression-11") takes uploads until 13 Oct and is scored by 20 Oct
Entry cost	0.25 τ registration (about $72), plus the miner's own model bill: each submission runs on the miner's own OpenRouter key
Scoring	Weighted tokens (input ×1, cached ×0.1, output ×3) against a baseline, with a penalty if the compressed agent solves under 80% of the tasks the baseline solves; tasks are SOMA's own hidden set. Winners are approved by hand
Entries and prize	56 uploads so far in round 11; prize pool 960 τ (41,328 α, about $278k at today's price)
Split, on chain	Three slots hold 75%, 15% and 10% of miner incentive; 37 more carry token weights. The 3 October leader cut weighted tokens about 64% while keeping at least 80% of the solve rate

Supply side

Miners paid on chain fell from 57 to 36 since 13 Jun; peak 214 on 5 Jul

Miner slots (UIDs) validators paid each day, read from chain (latest block 9,248,983); the owner's own key is left out because the chain burns its share. Slots, not companies. Since late July the contest pays per complexity element; on 11 Oct three slots held 99.9% of miner incentive (75%, 15%, 10%) and the rest carry the token screener weight of 0.002% each, so the count tracks the scoring rules more than the number of serious competitors.

The contest code is open; the product and the miners' submissions are not

What is open source, and what Dendrite keeps privateGitHub, 11 Oct 2026. V
Component	Public?	Detail
Contest platform, validator, benchmark	Yes, MIT	github.com/DendriteHQ/SOMA, SOMA-benchmark, SOMA-shared
Miner submissions	No	Held in a private store; no written licence governs them, so Dendrite's right to ship them rests on practice
OpenClaw compressor	Yes, MIT	Dendrite's own; credits no miner
The SOMA App proxy and its compressor	No	Which compressor serves customer traffic is not published

The owner holds 23% of alpha, locks a third of it, and has sold none

Owner and emission reads, chain, block 9,260,669 (11 Oct 2026)archive.chain.opentensor.ai: SubnetTAO, SubnetAlphaIn/Out, SubnetAlphaOutEmission, SubnetTaoInEmission, OwnerLock, OwnerCutAutoLockEnabled, StakeInfo for the owner coldkey. TAO $289.30, Kraken. V
Alpha price and market cap	0.02324 τ = $6.72; 1,989,868 α = $13.4M
TAO in the pool	9,843 τ ($2.8M), up from 8,841 τ on 2 Oct
Alpha issued	7,200 α/day = $48.4k/day, $17.7M/yr at spot, 132% of market cap
TAO the network injects (its cash cost)	48.4 τ/day = $14.0k/day; 4.2% of all subnet TAO inflow, ninth of 93 subnets
The owner coldkey's holding	457,613 α = 23.0% of circulating, $3.1M, up from 439,300 α on 2 Oct. The key has signed four transactions ever, so none of its alpha has been sold from it
Owner lock	146,040 α locked (7.3% of circulating, 32% of the owner's stack); automatic locking of the 18% cut is off
Owner's 18% cut	$8.7k/day, $3.2M/yr; against $0 of customer fees

No customer money reaches the token today

Where $1 of customer savings goes today, and what would have to change for any of it to reach the tokenMechanisms from the site and chain above; the last column is house judgment.
Party	Today	What would change it
Customer	Keeps the whole saving, about $0.10 per $1 of model bill on SOMA's own measure	A published fee
Dendrite	$0.00: "zero margin"	A share-of-savings or enterprise fee; compression vendors charge per token removed
Miners	Emission only	Unchanged unless revenue funds the pool
Alpha holders	$0.00: no buyback, burn or fee sink	A revenue-funded buyback; none announced

Dendrite's stated plans

SOMA's stated plans and their statusthesoma.ai and @SomaSubnet, read 11 Oct 2026. S
Plan	Source	Status
Measured benchmark release replacing the estimates	thesoma.ai	Announced, not published
Claude Code support; Zed	thesoma.ai	"Soon"; "Planned"
SOMA for Enterprise: private deployment, per-team budgets, savings report	thesoma.ai	"In design"
Spend part of emission on Lium compute and Say GM credits	@SomaSubnet, 28 Sep	Intent; supports the Gamma-token proposal (@DendriteHQ, 6 Oct)
Jev decision model for miners	@SomaSubnet, 7 Oct	Live for miners since 7 Oct

Bear, base and bull cases

Bear, base and bull cases over the next twelve months, with the evidence that would mark eachHouse judgment.
Bear	The proxy stays free and DeepSeek-only; OpenAI and Anthropic compaction makes it redundant on their models; engineering activity keeps falling; the token follows emission and root-basket flows
Base	Claude Code support and a measured benchmark of about 10% ship; an enterprise tier finds a handful of DeepSeek-heavy users; ecosystem traffic (Albedo, Gamma spending) grows; no customer money reaches alpha
Bull	A published fee on savings, a named outside customer, and a buyback funded from that fee; the contest keeps beating the published methods on new workloads

What would change our mind

A measured saving above 20% on real Copilot or Codex traffic, net of cache effects, would move the product case.
A published fee, or any customer, usage or revenue disclosure, would make SOMA a business with a price.
A buyback funded from that fee, above a quarter of revenue, would make the token a claim on the business rather than on emission.
The owner coldkey selling, or commit activity staying near September's 19 a month, would weaken the case.

Open items

Revenue was searched on the site, dashboard, docs, @SomaSubnet and @DendriteHQ (X, to 11 Oct), TAO Daily's six SOMA articles and taorevenue. The 11 Oct Novelty Search episode and Yuma's Subnet Spotlight were not reviewed.
Which compressor serves SOMA App traffic, and whether any miner's code is in it, is not published.
The "SOMA Lock Program" recorded on 16 Sep from a docs page title could not be found again; treat as unverified.