3code benches 75 lower token use vs opencode on 10-task SWE-Bench subset
I just ran the first formal validation of 3code!
Now this is a very limited benchmark- 10 tasks across 1 model- but the results are promising to say the least.
I used 10 tasks from SWE-bench as a preliminary test pool that have the same distribution of difficulty as the complete SWE-bench. I ran them using both 3code and opencode using the standard SWE-bench test runner, and a lightllm proxy configured in async mode to just watch and count the tokens used.
The results are remarkable- 3code used 75% less tokens across the board, and succeeded in a task opencode failed at. This is the exact same model- the most capable open weights model (with weights released, not just announced) at the time of writing, GLM 5.2, just a different coding agent.
I'd like to add that working on the system prompt and token savings wasn't a huge part of developing 3code just yet- and the greatest token saver for longer running tasks, 'cybernetic mode', isn't autoloaded in 3code just yet, so this is purely low hanging fruit common sense basic stuff- verbally compacting the system prompt and instructing to search isntead of reading everything, that sort of stuff, and having the prompt be modular. In other words, 3code is already getting by on a quarter of the tokens, and is just getting started.
Tokens were provided by z.ai startup program. Thank you very much!