A competition writes our compressor

One host serves the panel and the API. Point any OpenAI-compatible client at it with a SOMA key and requests are compressed on the way to the upstream model - the conversation and the chain-of-thought, never the instructions.

A credit is a cent, and it is passed through without a margin: what the upstream model costs is what the request costs. New accounts start with five dollars of credit, and top-ups take a card or TAO.

Subnet 114 runs competitions on Bittensor. Miners submit compressors, validators run them against real software-engineering tasks, and the score is the one that matters: what the task cost in tokens against whether it still passed.

A screener keeps the competition honest - two stages that a submission has to clear before its results count for anything. Everything the validators measured is public on the dashboard, task by task and run by run.

After a competition, our compression team takes what the miners wrote and does the part a competition cannot: making it work inside a live product, on real traffic, at the latency an agent will tolerate. That step is the reason the subnet and the app are one thing rather than two.

The subnet dashboard is public, with every competition's scores and token counts. The OpenClaw compressor is open source. The documentation covers how scoring, screening and the competitions actually work, including the parts that are awkward.

So is the accounting of what compression is worth: the site labels an estimate as an estimate, and says plainly where caching beats compression outright.