Trend questions now see every reading. get_history used to hand the model an even sample of at most 200 rows, one reading per 50 minutes across a week, so an hour-long spike contributed about one sample and a question like "why does the air quality spike every evening" could not be answered from the data with any output budget. Long ranges now come back as buckets that cover every reading (min, max and mean per interval for numeric sensors, dominant state and change count for on/off ones), sized so a week is hourly, at fewer tokens per sensor than the sample cost. Short ranges are still returned raw.
A cut-off answer says so, and long answers get room to finish. The default written ceiling goes from 2048 to 8192 tokens. That budget is shared with a reasoning model's hidden thinking, which grows with the prompt, so a long analytical answer needs more room than its visible length suggests. A ceiling is not a reservation, so short answers cost what they always did. MAX_OUTPUT_TOKENS, or the add-on's new Maximum answer length option, overrides it. A reply that ran out of room but had written something used to be delivered silently; both engines now end such a reply with a sentence saying so. Every cap hit is logged with the token numbers.
A question typed into Assist is not a spoken one. The integration decided "spoken" from agent_id, which modern Home Assistant sets on every routed request, so every question ran under the short spoken ceiling. It now keys on satellite_id, set only for voice satellites. If you installed the integration from HACS, update it to 0.10.2 for this part; the add-on in this repo is pinned to this version.
All three carried over from Nives 2.5.8 and 2.5.10, where they were found on a live house.