Explainers · code agents
10 code agents, one Qwen 3.8 27B: 4x spread in tokens, zero difference in pass rate
On October 30, 2026, the second part of the code agent test on a local llama-server with Qwen 3.8 27B was published on Reddit. All agents passed 5 of 5 tasks, but median token consumption differs by 4.2x. We break down the mechanics: why pass rate stops distinguishing agents and what cache reuse actually shows.
30 Oct 2026
7 min read
