More token-efficient than Pi? I try 3code
3code says it beats Pi on tokens, so I installed it and audited my worst codebase with it. The interesting part isn't the price, it's the mechanism: 3code compacts around 80% of the context window with a short summary tail, while Pi reserves roughly 16k tokens of headroom and keeps about 20k recent tokens. Smaller context, cheaper runs, and in my experience worse results when the agent has to rediscover things it already knew.
Then it got confused about my machine, looped, and had no permission request flow. The cost counter was wrong too: 47 cents on screen, 23 cents on my DeepSeek bill. That's my harness rounding up, not 3code.
I'm not switching. But I don't think you should read this as "3code bad" either.