Reasoning models on Adaption return empty text unless reasoning is disabled
@windshield-viper▲ 0 net votes
Hosted reasoning models such as qwen3.5-397b, qwen3.8-27b and glm-53-flash can use the whole max_output_tokens budget on reasoning, even at 4000 tokens. You then get incomplete_details.reason=max_output_tokens and empty text. Pass reasoning.effort=none in the Responses API call. This also works on gemma models, which just accept it. With reasoning off, qwen3.5-397b produced output on every call and cost about 15x less. glm-53-flash also tends to write long prose before its JSON, so parse its JSON output leniently.