Reasonix: a coding agent built around the bill, not the model
Reasonix is a terminal coding agent that only works with DeepSeek, and that constraint is the entire point. Most coding agents treat the model as swappable, plug in Claude, GPT, whatever you fancy. esengine went the opposite way. Every layer of Reasonix is tuned to one specific DeepSeek mechanic, the byte-stable prefix cache. Keep the prompt prefix identical byte for byte and DeepSeek bills those cached tokens at a fraction of the normal price. So Reasonix is engineered to never break that prefix. Their north star, in their own words, is a coding agent cheap enough to leave running.
That is a very different bet from everyone else in the race. The whole field is sprinting to make agents smarter, better planning, more tools, bigger context. Reasonix bets the bottleneck was never intelligence, it is cost. If an agent is cheap enough you stop rationing it, you just leave it on, and leaving it on quietly changes how you work. It hit the Hacker News front page today at 350 plus points and 5.5k stars, and the native Tauri client puts live cost, cache and token meters right at the bottom of the window. They literally put the bill on screen.
The part worth chewing on is what this says about DeepSeek itself. The prefix cache is cheap and predictable enough that one person can build an entire product on the economics alone, not the smarts. You install it from npm, run reasonix code with a DeepSeek API key, and that is it. Opinionated, not general, every abstraction justified by a DeepSeek-specific behavior.
The takeaway I keep coming back to. We have spent two years asking whether the model is good enough. Reasonix asks the question almost nobody is building around: is it cheap enough to never turn off. Repo is at https://github.com/esengine/DeepSeek-Reasonix
← Back to all articles
That is a very different bet from everyone else in the race. The whole field is sprinting to make agents smarter, better planning, more tools, bigger context. Reasonix bets the bottleneck was never intelligence, it is cost. If an agent is cheap enough you stop rationing it, you just leave it on, and leaving it on quietly changes how you work. It hit the Hacker News front page today at 350 plus points and 5.5k stars, and the native Tauri client puts live cost, cache and token meters right at the bottom of the window. They literally put the bill on screen.
The part worth chewing on is what this says about DeepSeek itself. The prefix cache is cheap and predictable enough that one person can build an entire product on the economics alone, not the smarts. You install it from npm, run reasonix code with a DeepSeek API key, and that is it. Opinionated, not general, every abstraction justified by a DeepSeek-specific behavior.
The takeaway I keep coming back to. We have spent two years asking whether the model is good enough. Reasonix asks the question almost nobody is building around: is it cheap enough to never turn off. Repo is at https://github.com/esengine/DeepSeek-Reasonix
Comments