Add test results for GLM 4.5 (thinking & non-thinking modes) model to polyglot leaderboard

[Oct4Pie]

The 2 commits add the evaluation results for the GLM 4.5 model in both thinking and non-thinking modes.

The tests were done in diff editing format. There were many errors in tool output formatting in thinking mode (due to poor structured output adherence) but it should not affect the test results as those were retried but the costs reported are not accurate.

model settings used for non-thinking mode:

- name: openrouter/z-ai/glm-4.5
  extra_params:
    extra_body:
      reasoning:
        enabled: false
      provider:
        ignore:
          - novita
    max_tokens: 96000

for thinking mode, reasoning.enabled was set to true.

Just a note: provider novita was excluded from the openrouter providers for reproducibility as it did not explicitly mention the quantization of the model. All other providers used the fp8 variant.

The benchmark exercise folders are attached for reference

2025-08-03-11-33-59--glm-4.5-thinking-polyglot.zip
2025-08-03-13-07-25--glm-4.5-polyglot.zip

[kneelesh48]

Please use the official api as providers on openrouter often quantize the model.

[nuireprog]

It work on endpoint API openai

model: openai/glm-4.6

On redirige la requête vers le endpoint de z.ai

openai-api-base: "https://api.z.ai/api/coding/paas/v4"

REMPLACEZ CECI par votre vraie clé API obtenue sur https://z.ai/

openai-api-key: "YOURAPIKEY"

[Kreijstal]

what are the results for glm-4.6?