GLM 5.2 edges MiniMax M3 on correctness, but costs 2.8× more
A head-to-head autonomous coding benchmark pitting GLM 5.2 against MiniMax M3 across 60 hidden-graded tasks found GLM more accurate (92% full-pass vs. 84%) but nearly three times more expensive and almost twice as slow.
Score breakdown
The benchmark shows that for autonomous coding agents, the choice between GLM 5.2 and MiniMax M3 reduces to a concrete cost-accuracy tradeoff: GLM's correctness edge is real but narrow and concentrated in greenfield packaging, while MiniMax delivers nearly the same results on modification tasks at roughly one-third the cost and half the latency.
- 01GLM 5.2 scored 92% full-pass (165/180) with a 0.976 mean score across 60 hidden-graded tasks
- 02MiniMax M3 scored 84% full-pass (152/180) with a 0.961 mean score
- 03MiniMax M3 cost $6.67 for the scored runs vs. GLM 5.2's $18.47
Brandon Huey used Thinkbench — a custom evaluation harness — to drive both GLM 5.2 and MiniMax M3 through the same autonomous coding loop: reading files, writing files, running shell commands, and stopping when the task was complete. The scored suite comprised 60 tasks across greenfield builds, bug fixes, feature additions, and repair-to-green scenarios, run 3 trials per model for 180 scored rows each. Hidden graders applied fixed-denominator behavior checks after each model stopped. A separate set of 12 ungraded tasks used deliberately vague briefs — covering systems like audit logs, schedulers, feature flags, and notification hubs — to observe how each model handles ambiguous instructions.
GLM 5.2 came out ahead on correctness: 165/180 full-pass (92%) and a 0.976 mean score, versus MiniMax M3's 152/180 (84%) and 0.961 mean score.
GLM 5.2 came out ahead on correctness: 165/180 full-pass (92%) and a 0.976 mean score, versus MiniMax M3's 152/180 (84%) and 0.961 mean score. MiniMax M3 used more tokens on average (135,060 vs. 82,443) but cost significantly less ($6.67 vs. $18.47) and was faster (45s vs. 80s average latency). On modification tasks — bug fixes, feature additions, repair-to-green — both models effectively saturated the benchmark at 0.999–1.000 mean score. The real separation appeared in greenfield builds: 54 of 60 scored tasks landed within 0.1 mean score of each other, and all six larger gaps came from from-scratch tasks. GLM's biggest win was `ticketflow` (GLM: 1.00, MiniMax: 0.33 average), where MiniMax's failures stemmed from package layout issues that prevented the hidden grader from importing from the workspace root. MiniMax's clearest wins were `patchwise` (MiniMax: 1.00, GLM: 0.62) and `migrato`, where GLM lost points to implementation bugs including a name typo and trailing-newline diff handling errors. On the ungraded ambiguity tasks, MiniMax M3 consistently added more production-shaped machinery when prompts left room, while GLM 5.2 tended to stay closer to the plain reading of the brief.
Key facts
- 01GLM 5.2 scored 92% full-pass (165/180) with a 0.976 mean score across 60 hidden-graded tasks
- 02MiniMax M3 scored 84% full-pass (152/180) with a 0.961 mean score
- 03MiniMax M3 cost $6.67 for the scored runs vs. GLM 5.2's $18.47
- 04MiniMax M3 averaged 45 seconds per run vs. GLM 5.2's 80 seconds
- 05On modification tasks (bug fixes, feature additions, repair-to-green), both models scored 0.999–1.000 mean — nearly identical
- 06All six largest task-level gaps were concentrated in greenfield builds
- 07On ambiguous-brief tasks, MiniMax M3 added more production-shaped machinery while GLM 5.2 stayed closer to the plain brief
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 20, 2026 · 08:55 UTC. How this works →