Hi, While not GLM per se, I find it that setting it through providerOptions works better for most of the models that I used. You can find the reasoning config for each model from their official documentation (For GLM: https://docs.z.ai/guides/capabilities/thinking-mode ) and perhaps need to check if different providers have different config (e.g. Bedrock).

Overview - Z.AI DEVELOPER DOCUMENT

Thinking Mode - Overview - Z.AI DEVELOPER DOCUMENT

1
"providerOptions": {
2
    "gateway": {
3
        "only": ["zai", "novita"],
4
        "sort": "ttft"
5
    },
6
    "zai": {"thinking": {"type": "disabled"}},
7
    "novita": {"thinking": {"type": "disabled"}}
8
}

Though it does that mean you need to set them for each providers that you use.

Hi Hugopod,

Your table makes this look less like a client-side syntax issue and more like provider-specific behavior behind the Gateway.

One thing I’d do for the benchmark is pin the provider instead of letting Gateway route across all providers, since you already found that Baseten is the only one honoring the disable option:

1
providerOptions: {
2
  gateway: {
3
    only: ["baseten"]
4
  },
5
  baseten: {
6
    thinking: { type: "disabled" }
7
  }
8
}
9

Then run the same prompt a few times and compare:

1
- first token latency
2
- total output tokens
3
- reasoning_tokens
4
- provider metadata / selected provider
5

If Baseten consistently returns reasoning_tokens: 0 and the others do not, then I’d avoid fallback routing for this latency-critical path unless each fallback provider has its own verified native “thinking disabled” option.

I’d also be careful with reasoning: { effort: "none" } as a portable assumption here. In the AI SDK docs, none is only supported for specific OpenAI GPT-5.1 models, so for GLM via multiple providers I’d expect provider-specific options to be more reliable than a single normalized field.

For this specific case, the strongest support report would be exactly what you already collected: same model, same prompt, same Gateway request shape, pinned provider, and reasoning_tokens still generated despite a disable option.