69 trillion tokens. in one week. that's OpenRouter's global LLM traffic for Aug 3–9, up 21% week over week.
and the top four? all open-weight models. DeepSeek-V4-Flash leads at 8.83T, then Tencent's Hy3, DeepSeek preview, Xiaomi's MiMo-V2.5. open-weight has now out-ranked US labs for 15 straight weeks.
the kicker: agent + coding use is just 19% of calls, but burns 81% of all tokens. people aren't chatting — they're running machines.
so is this a quality win, or just the cheap option winning by default?
because if the best-value models are all open-weight, the "frontier" gap everyone fears might already be a pricing gap.
quality lead, or just cheaper?
Top comments (0)