Using a proprietary evaluation harness called Thinkbench, both models were tasked with executing the same autonomous coding loops. The evaluation covered 60 tasks, including file read/write operations and shell command execution.
In the scored tasks, GLM 5.2 achieved a full pass rate of 92% with an average score of 0.976. Meanwhile, MiniMax M3 had a full pass rate of 84% and an average score of 0.961. While the difference is slight, cost and latency must also be considered.
Regarding costs, MiniMax cost $6.67 compared to $18.47 for GLM. Average execution times also differed significantly, with MiniMax averaging 45 seconds and GLM averaging 80 seconds.
There was almost no difference between the two models when handling existing code. For tasks such as bug fixes and feature additions, average scores converged between 0.999 and 1.000. Clear differences emerged only in tasks requiring construction from an empty repository.
Out of the 60 tasks, 54 showed an average score difference of 0.1 or less. The six tasks with significant differences were all zero-base construction tasks. For example, in ticketflow, GLM scored 1.00 while MiniMax scored 0.33; this was because the grader could not load MiniMax's package configuration.
Conversely, in patchwise, MiniMax scored 1.00 while GLM scored 0.62. GLM produced bugs related to naming typos and newline processing.
Response to ambiguous instructions was also compared. MiniMax proactively added production-ready mechanisms, such as hash chain verification and query builders. GLM tended to choose flatter implementations that more literally followed the instructions.
Based on an analysis of the generated output, MiniMax showed a stronger inclination toward implementing locks, persistence, and fallback paths.
Source: MiniMax M3 vs. GLM 5.2: Codegen comparison across autonomous coding tasks (HN 55pt, 19 comments) (HN Search (backfill), 2026-06-20)