MiniMax M2 vs GLM 4.6 vs Claude Sonnet 4.5: A Hands‑On Programming Model Comparison

Last month the author compared two large models; readers asked for more, so the author added MiniMax’s newly released M2 model and tested it alongside the Chinese GLM 4.6 and Anthropic’s Claude Sonnet 4.5, all of which are among the most advanced programming‑oriented LLMs.

Background

MiniMax released M2 on October 27, positioning it as the latest domestic flagship model. A HuggingFace community lead reported that M2 ranked fifth worldwide in the Artificial Analysis benchmark and first among open‑source models, while OpenRouter placed it third in global call volume.

Test Environment

The author used the official web‑based agents (MiniMax Agent for M2, domestic and overseas versions) which are free for a two‑week trial; API pricing after the trial is 2.1 CNY/8.4 CNY per million tokens, about 8 % of Claude’s price.

All three models were accessed through their respective web interfaces, and the author followed the same prompts for each test.

Task 1 – Code Execution

The author cloned Simon Willison’s llm repository ( https://github.com/simonw/llm) and ran the test suite:

pip install -e '.[test]'
pytest

MiniMax Agent executed the commands in a sandboxed environment, completing the run in about three minutes and reporting that all 466 tests passed. It also returned a coverage analysis highlighting which code paths were exercised—an auxiliary output the author had not requested and had not seen from the other models.

Task 2 – Code Generation & Database Modification

The author asked the model to extend the same repository by adding a new feature that required code changes, a database schema alteration, and new pytest cases. The prompt specified four steps: (1) describe the repository, (2) add a parentresponseid column to model a tree‑structured conversation, (3) write new tests, and (4) create a tree_notes.md file.

During execution MiniMax automatically fetched the repository via deepwiki.com and later consulted datasette.io for SQLite analysis—behaviors the author did not anticipate. After completing the task, MiniMax produced a detailed summary of the modifications, added an example file, and displayed a diagram of the updated conversation structure, again without being prompted to do so.

Task 3 – Creative Reasoning (Pelican on a Bicycle)

The final test asked the model to generate an SVG image of a pelican riding a bicycle, a scenario that does not exist in reality and requires the model to reason creatively. MiniMax produced an SVG that included a road and a reasonably correct bicycle frame, though it missed the handlebars and the pelican’s pose could be improved.

For comparison, the author also displayed the SVGs generated by GLM 4.6 and Claude Sonnet 4.5.

Observations & Conclusions

MiniMax M2 consistently delivered correct functional results, added helpful auxiliary information (coverage reports, design summaries, extra example files), and handled the creative SVG task with a more accurate structural output than the other models. The author notes that the model’s user‑friendly approach—providing explanations and extra context—enhances perceived reliability.

Given the strong performance, free trial period, and competitive pricing, the author plans to adopt MiniMax M2 in future work and recommends others try it.

(End)