Your agents pay for tokens they never needed.
Every turn, a coding agent re-sends the whole conversation - files it already read, tool output it already used, its own reasoning. You pay for all of it. SOMA compresses the context and the chain-of-thought on the way to the model, and keeps what the task needs.
Context grows every turn. So does the bill.
- AAgents re-send history on every request - the same files, the same tool output.
- BCost per task grows faster than the task itself: every turn pays for all the turns before it.
- CMost of that context is no longer needed to get the next step right.
At turn 14
This session has sent 658k tokens so far. 39k of the last request was context the model no longer needed - 45% of it.
Tokens sent per agent turn
Illustration - a typical 14-turn coding session. Real figures land with the benchmark release.
SOMA
We build the layer that makes AI cost what it should.
SOMA is a research company working on one problem: how little context a model needs to do the job well. The app is where that work ships first - it isn't the only place.
SOMA App
Point your coding agent at SOMA instead of the model. We compress the conversation and the reasoning on the way through, and charge the model's real cost - no markup.
app.thesoma.aiSomarizer
Paste text or drop a PDF, pick how hard to compress, get a shorter version back. The same compression, without writing any code.
thesoma.ai/somarizerCompressor for OpenClaw
A context compressor that plugs into the OpenClaw agent loop. Public repository, run it yourself.
github.comSOMA for Enterprise
Private deployment inside your cloud, per-team budgets and a savings report your finance team can read. In design - talk to us if you need it.
Talk to usMeasured, not claimed
A compressor only counts if the task still passes. We score both, and publish the runs.
Zero margin
You pay what the model costs us. We earn when compression earns its keep, not on the spread.
Written by competition
Our compressor comes from an open subnet where anyone can beat the current one.
Benchmarks
Same agent. Same model. Fewer tokens.
We run the same agent on the same task twice - with compression off and on - and compare cost and whether the task still passed. Today every figure below is our estimate; measured runs replace them as they land, and each figure says which it is.
List price against the SOMA price for the same input. Every pair is our current estimate; measured runs replace them as they land.
Estimated savings - today
30% - GitHub Copilot CLI + DeepSeek V4 Pro, for this pair. SOMA labels every figure as measured or estimated - and never presents one as the other.
What we test on
- 1Real issues from real GitHub repositories
- 2Navigating large codebases and changing many files
- 3Debugging, running and fixing tests
- 4Long tasks with many tool calls and reasoning rounds
Savings depend on the task, the context size and how much of it the provider already caches. Long sessions with heavy tool output save the most; short ones can save nothing. List prices move and differ by where a model is served - the app shows the current ones. In your dashboard token counts are measured; the dollar figure is an estimate.
Agents & models
Keep your agent. Keep your model.
SOMA speaks the OpenAI API, so the agent doesn't have to change. Point it at our endpoint and it starts sending less.
Your agent
The model
Setup
One line, then keep working.
Sign up, pick your agent and a model, copy your sk-soma key, run one command. The agent you already use keeps working, just with less going out.
# point Copilot CLI at SOMA with your key and modelcurl -s https://app.thesoma.ai/copilot \| bash -s sk-soma-YOUR-KEY deepseek/deepseek-v4-pro# that's it - start codingcopilot
macOS / Linux. On Windows: run it in WSL or Git Bash, not PowerShell.
Compression runs before the request leaves us. If the compressor is unavailable, your request goes through uncompressed rather than failing.
Research & development · Bittensor subnet 114
Our compressors come from an open competition.
SOMA runs subnet 114 on Bittensor. Miners compete to build the best compressor, validators score every submission on real work, and after each competition our compression team takes the winning code and adapts it to the app.
- 01
Miners submit
Independent teams train and submit compressors to the subnet.
- 02
Validators score
Every submission runs on real tasks. Cost and task success, both.
- 03
Our team adapts it
After each competition, our compression team takes the winning code and makes it production-ready.
- 04
It ships to the app
The adapted compressor runs in front of your agent, and usage funds the next round.
- Subnet
- SN 114
- Compresses
- Context + CoT
- Miners last round
- 162
- Leaderboard
- public
Get started