An OpenAI flagship lost 38% of its daily usage in three days - then set three straight all-time highs

Everyone publishes AI model leaderboards. Almost nobody publishes the day-level curve - and the curve is where the actual story is. At whatstrending.ai we snapshot every model's daily request count from OpenRouter's usage feed into a database, once a day, INSERT OR IGNORE - a row can never be revised after the fact. We're 78 days in. Here is what openai/gpt-5.6-luna did over the last two weeks: | Date | Requests/day | |---|---| | Aug 04 | 162.9M | | Aug 05 | 180.8M | | Aug 06 | 193.8M | | Aug 07 | 210.8M | | Aug 08 | 222.5M ← peak | | Aug 09 | 209.6M | | Aug 10 | 171.6M | | Aug 11 | 137.3M ← floor: −38.3% off peak in 3 days | | Aug 12 | 143.2M | | Aug 13 | 157.3M | | Aug 14 | 231.8M ← new all-time high, +47.4% in one day | | Aug 15 | 237.3M | | Aug 16 | 243.5M ← third consecutive all-time high | A flagship model shed more than a third of its daily traffic in three days, then won all of it back - and more - inside the same week. Three consecutive days above the old peak as of Aug 16. Why this isn't a data artifact Three checks, run before writing this: - The whole table didn't move. The all-model daily total went 3.56B → 3.66B (+2.8%) on the day luna jumped 47%, and 3.70B → 3.73B (+0.8%) on Aug 16. Luna's share moved: roughly 4.4% → 6.5%. - Peers didn't dip with it. During luna's drawdown (Aug 09-11), deepseek/deepseek-v4-flash-0731 andtencent/hy3 were setting their own highs on the same days - so this is not a provider-wide trailing-edge reporting droop. - Rows are frozen at write time. Aug 14 read 231.8M on the day we collected it and reads 231.8M today. Nothing was backfilled or revised. Why it matters Monthly aggregates - which is what every public ranking shows - would have flattened this entire event into "usage grew". The day-level series shows real flagship usage is far more volatile than anything the leaderboards suggest: a −38% drawdown and a full recovery to new highs fit inside eight days. If you're making routing, capacity, or "which model is winning" decisions off monthly numbers, you're averaging away the actual signal. The live, daily-updated version of this curve is here: https://whatstrending.ai/models Method: daily snapshot of OpenRouter per-day request counts, 100 models/day, 78 days and counting, stored append-only (INSERT OR IGNORE ) in Cloudflare D1. Top comments (0)