SOMA trims what a coding agent re-sends to its model; the team measured about 10% on Copilot. It bills model cost plus 30% of the saving and says revenue began after launch, amount undisclosed (fund call, 15 Oct).
| Claim | Who said it | What the check shows |
|---|---|---|
| SOMA cuts an agent's token cost about 15% | thesoma.ai benchmark table | Every row is labelled "est.". The compression page puts the measured Copilot figure "far lower", about 10%, and promises "real figures" with the benchmark release |
| "Around 10% token savings" at the Copilot launch | @SomaSubnet, 28 Aug | Matches the compression page; no per-customer or aggregate measurement is published |
| One prompt cut from 12,937 to 7,801 tokens, −39.7% input cost | TAO Daily, 7–13 Sep (single source) | One team-supplied prompt, not reproduced; token counts overstate the saving where a cache serves the prompt |
| "Our compressor comes from an open subnet where anyone can beat the current one" | thesoma.ai | Nothing published ties a winning submission to the production proxy; the only released compressor credits no miner |
| Integrated with Chutes, Say GM and Engy; proxying traffic to Albedo | @SomaSubnet, 9 Oct | Integrations, not paying contracts; Albedo (SN97) is another Dendrite subnet; volumes not disclosed |
| Item | Detail |
|---|---|
| SOMA App (open beta) | A proxy for GitHub Copilot CLI and Codex that trims history, tool output and reasoning; Claude Code "soon", Zed "planned" |
| Models | Three DeepSeek models only; no OpenAI or Anthropic |
| Price | "Zero margin. You pay what the model costs us." DeepSeek V4 Pro input $0.92 per million tokens, $0.78 "with SOMA" (est.) |
| How Dendrite earns | "We earn when compression earns its keep, not on the spread"; 30% of the saving, about 9% of the baseline bill, from prepaid credits (Aleksander Muszyński, fund call, 15 Oct) |
| Customers | None named; revenue claimed on the fund call (15 Oct), no figure anywhere |
| Item | Detail |
|---|---|
| Company | Dendrite Quantum LTD, Cyprus (HE 475409); also builds Teutonic (SN3) and Albedo (SN97); sister company BlockWise, Warsaw |
| Founder | Piotr Barbachowski, Limassol; calls Dendrite 50 engineers, grown from five who first mined TAO in September 2022 |
| Named staff | Aleksander Muszyński (integrations), Jan Różycki and Hubert Korzeniewski (miner workshop); on Discord "Matt | SOMA" (top committer) posts scoring changes, "oli | SOMA" holds the owner role |
| Team size, checked | Eight people have ever committed to SOMA's five public repositories, three in October; the 50-engineer figure covers three subnets and BlockWise |
| Method | Who | Saving | Effect on tasks solved | Cost to use |
|---|---|---|---|---|
| SOMA App | Dendrite | 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 |
The methods that keep task success intact are simple: drop old tool output and prune file reads. The contest leader cut weighted tokens about 64% on SOMA's hidden tasks (3 Oct; SOMA's scoring, unreproduced); the product has not shown that gain on real traffic.
| Alternative | What the buyer gets | Price |
|---|---|---|
| Provider prompt caching | Anthropic cache hits $0.25 against $10 per million input tokens (docs); DeepSeek $0.003 against $0.15 (pricing) | Included; cache writes 1.25–2× list at Anthropic |
| Harness auto-compaction | Claude Code, Codex, Gemini CLI, Cursor and OpenClaw prune old turns themselves | Included; the summary is billed as tokens |
| Gateways | OpenRouter routes at provider price and keeps cache hits sticky (docs); LiteLLM, Portkey and Kong bundle open-source compressors | No token markup; enterprise tiers |
| Compression start-ups | Compresr (compression API), The Token Company (YC) | Usage-priced; no tier above $0.30 per million tokens (2 Oct) |
| Measure | Figure |
|---|---|
| Floor: what a compression vendor charges today | No standalone vendor discloses revenue; public prices 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 input not already served from cache, roughly 5–35% of input cost on a cache-heavy harness; a 30% cut there is about 10% of the bill (house arithmetic) |
| Companies doing this in-house | Cursor $29.3B, Cognition $10.2B, Factory $5B, OpenRouter sold to Stripe for over $7B (single source); for all, compression is a feature, not a product |
| Who | What they do | How they are paid | Per day | Customer money? |
|---|---|---|---|---|
| Miners (three paid slots) | Submit code that rewrites the agent's message list, scored on tokens saved on hidden tasks | New alpha; 75%, 15% and 10% of incentive | $19.8k | No |
| Validators and stakers (nine) | Set weights from SOMA's open-source scoring and pay delegators the staking yield; the owner's validator takes 55% of dividends | New alpha | $19.8k | No |
| Dendrite, the owner (coldkey 5FHrQ…) | Runs the contest, the task set and the proxy | The 18% owner cut in alpha | $8.7k | No |
| Alpha holders | No buyback, burn or fee sink; taorevenue, 30 days: owner inflow $0, burn $0, outflow $1,138,744 | — | — | No |
| Question | Answer |
|---|---|
| Cadence | Two weeks a round: one of uploads, one of evaluation; round 11 takes uploads until 13 Oct, scored by 20 Oct |
| Entry cost | 0.25 τ registration (about $72) plus the miner's own model bill on its own OpenRouter key |
| Scoring | Weighted tokens (input ×1, cached ×0.1, output ×3) against a baseline, penalised below 80% of the baseline's solve rate; hidden tasks; winners approved by hand |
| Entries and prize | 56 uploads in round 11; prize pool 960 τ (41,328 α, about $278k) |
| Split, on chain | Three slots hold 75%, 15% and 10% of miner incentive; 37 more carry token weights |
UIDs paid each day, from chain (block 9,277,782); the owner's key is excluded because the chain burns its share. Three slots hold 99.9% of incentive (75/15/10); the rest carry a 0.002% screener weight, so the count tracks scoring rules, not competitors.
| Component | Public? | Detail |
|---|---|---|
| Contest platform, validator, benchmark | Yes, MIT | github.com/DendriteHQ/SOMA, SOMA-benchmark, SOMA-shared |
| Miner submissions | No | Private store; no written licence, 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 |
| TAO in the pool | 9,843 τ ($2.8M), from 8,841 τ on 2 Oct |
| TAO the network injects (its cash cost) | 48.4 τ/day = $14.0k/day; 4.2% of all subnet inflow, ninth of 93 subnets |
| The owner coldkey's holding | 457,613 α = 23.0% of circulating, $3.1M, from 439,300 α on 2 Oct; four transactions ever signed, so none sold from it |
| Owner lock | 146,040 α locked (32% of the owner's stack); automatic locking of the 18% cut is off |
| Party | Today | What would change it |
|---|---|---|
| Customer | Keeps the whole saving, about $0.10 per $1 of model bill | A published fee |
| Dendrite | 30% of each saving; amount undisclosed (fund call, 15 Oct) | An enterprise fee |
| Miners | Emission only | Revenue funding the pool |
| Alpha holders | $0.00: no buyback, burn or fee sink | A revenue-funded buyback; "modeled and prepared", timing open (fund call, 15 Oct) |
| Plan | Source | Status |
|---|---|---|
| Measured benchmark replacing the estimates | thesoma.ai | Announced, not published |
| Claude Code support; Zed | thesoma.ai | "Soon"; "Planned" |
| SOMA for Enterprise: private deployment, per-team budgets | thesoma.ai | "In design" |
| Spend part of emission on Lium compute and Say GM credits | @SomaSubnet, 28 Sep | Intent; backs the Gamma-token proposal (@DendriteHQ, 6 Oct) |
| Buyback from revenue | Fund call, 15 Oct | "Modeled and prepared"; timing open |
| GPT and Anthropic models through the proxy | Fund call, 15 Oct | "Pretty soon" |
| Jev decision model for miners | @SomaSubnet, 7 Oct | Live since 7 Oct |
| Bear | The proxy stays free and DeepSeek-only; provider compaction makes it redundant; commits keep falling; the token follows emission and root flows |
| Base | Claude Code support and a measured benchmark near 10% ship; an enterprise tier finds a few DeepSeek-heavy users; no customer money reaches alpha |
| Bull | Revenue at the 30% fee disclosed, a named outside customer and a buyback funded from that fee; the contest keeps beating published methods |