Assistants: 'the answer carried no text' when a thinking model spends max_tokens issue #201
closed: completed andymccall opened this · 0 comments
andymccall
doomsdayone/bomberman94 proposal 38: Claude (claude-opus-5, anthropic-messages) did not answer: the answer carried no text, after 54 s. Opus 5 thinks by default (adaptive, effort high) and returns empty thinking blocks; assistants::ask sends max_tokens 4096 with a 60 s timeout and never reads stop_reason, so a large proposal can spend every token thinking (stop_reason max_tokens, no text block), and a refusal (stop_reason refusal, stop_details.category) looks the same. Fix: raise max_tokens (about 16000) and stream so the timeout is not hit, or set output_config.effort per task (medium for summaries and drafts); read stop_reason and say why (cut off while thinking; declined: category); keep a longer timeout for streaming. Same check for chat-completions finish_reason length/content_filter.
doomsdayone/bomberman94 proposal 38: Claude (claude-opus-5, anthropic-messages) did not answer: the answer carried no text, after 54 s. Opus 5 thinks by default (adaptive, effort high) and returns empty thinking blocks; assistants::ask sends max_tokens 4096 with a 60 s timeout and never reads stop_reason, so a large proposal can spend every token thinking (stop_reason max_tokens, no text block), and a refusal (stop_reason refusal, stop_details.category) looks the same. Fix: raise max_tokens (about 16000) and stream so the timeout is not hit, or set output_config.effort per task (medium for summaries and drafts); read stop_reason and say why (cut off while thinking; declined: category); keep a longer timeout for streaming. Same check for chat-completions finish_reason length/content_filter.