3code Achieves 75% Token Efficiency Gain Over OpenCode in SWE-Bench Benchmark
Key Takeaways
- ▸3code achieves 75% token reduction compared to OpenCode on SWE-Bench tasks using GLM 5.2
- ▸Efficiency gains derive from basic system prompt optimization and modular architecture—advanced features like cybernetic mode remain disabled
- ▸3code succeeded on a task where OpenCode failed, indicating improved problem-solving capability alongside reduced resource consumption
Summary
Capocasa has published the first formal validation of 3code, a novel coding agent framework designed for software engineering tasks. In a preliminary benchmark using 10 representative tasks from SWE-Bench—maintaining the same difficulty distribution as the full suite—3code demonstrated remarkable efficiency improvements over OpenCode, consuming 75% fewer tokens while achieving equal or better task completion. The benchmark used GLM 5.2, the most capable openly-released large language model at the time of testing, ensuring a fair comparison between agent architectures rather than models.
The token savings appear to stem from fundamental optimizations that represent early-stage work. The creator notes that system prompt compression and architectural decisions favoring search-over-read comprise only the initial phase of optimization for 3code. Critically, 'cybernetic mode'—described as the greatest token-saving feature for longer-running tasks—remains disabled in the current release. This suggests the observed 75% efficiency gain represents basic architectural improvements with substantial optimization potential remaining.
While the benchmark scope is limited to 10 SWE-Bench tasks and a single model, results indicate significant potential. Notably, 3code succeeded on a task where OpenCode failed, demonstrating improved capability alongside efficiency gains. The research was supported by z.ai's startup program, which provided computational resources for testing.
- Validation is limited to 10 representative tasks; broader benchmark testing is needed to validate generalizability
Editorial Opinion
The 75% token efficiency improvement is striking precisely because it stems from straightforward architectural and prompt-engineering choices rather than novel techniques. This suggests considerable untapped optimization potential in the coding agent space, and 3code's modular design may represent a meaningful step toward resource-efficient AI-assisted development. However, the narrow benchmark scope demands expanded validation across a full SWE-Bench run and additional models before claiming general superiority.
