DeepSeek reroutes every V4-Pro API request to V4.1-Flash from 14 September

The $0.15 in DeepSeek's launch example is a cache price. The example, reported by VentureBeat, takes an agent with a 500,000-token reusable prefix hitting cache across 100 requests, so 50M cached input tokens, and prices it at about $0.15 off-peak against $25 on Claude Opus 5 ^[11]. Cached input bills at $0.003 per 1M tokens off-peak; uncached input bills at $0.15 ^[4]. Run the same 50M tokens uncached and the line item is $7.50, fifty times higher ^[1]. The launch material says the advantage narrows to the uncached rate when the agent cannot reuse a prefix ^[21].

The same example puts Kimi K3 at $15 for that traffic ^[11], which works out to $0.30 per 1M cached tokens, a hundred times DeepSeek's rate ^[7]. That matches the description of K3's rate card at roughly a hundred times the cached rate and twenty times the uncached input price ^[18].

DeepSeek produced every one of these scores itself ^[7]. DeepSWE v1.1 has V4.1-Flash at 74.2, Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0 ^[6], a margin of 0.2 points on an eval the vendor ran ^[4]. For that margin to say anything about your repository, your tasks would have to resemble DeepSWE's. The reasoning columns are wide enough to survive first-party grading: 36.8 against 56.3 on Humanity's Last Exam without tools, a gap of 19.5 points ^[8][5], and 20.3 against 37.0 on ProgramBench ^[9].

V4-Pro billed at $0.66 off-peak input and $1.98 off-peak output before the reroute ^[15]. At Flash's off-peak rates the same traffic costs 77% less on input and 70% less on output ^[2], and Bloomberg Intelligence put the effective saving at up to 32%, reversing an August increase ^[16]. I would weigh the Terminal-Bench 3.0 figure more heavily than the price move: Flash scores 30.0 there, ahead of the earlier V4-Pro checkpoint and behind both US models ^[10]. DeepSeek has not announced a date for a V4.1-Pro ^[3]. Whether a caller can pin the old checkpoint is unknown.

Peak hours are weekdays 01:00 to 04:00 UTC and 06:00 to 10:00 UTC ^[5], seven hours a day ^[3], and every rate doubles inside them ^[4]. Unattended batch work moved outside those windows costs half.

Self-hosting means holding 552B total parameters, up 94% from V4-Flash's 284B ^[17][6], while about 8B activate during prefill ^[17]. The layout is a new causal encoder-decoder with 40 layers split 20 and 20 ^[17].

DeepSeek's launch comparison gives the model 88.1 on CyberGym, the best figure in that table ^[14], and its technical report says it trails the best closed systems at interpreting complicated images ^[13]. The dev.to review that collected these numbers recommends V4.1-Flash for high-volume agent loops, Claude Opus 5 or GPT-5.6 Sol when a single hard answer matters more than the bill, and Kimi K3 when you need stronger reasoning but must keep open weights ^[20]. K3 scores 69.0 on DeepSWE v1.1 and leads on Humanity's Last Exam, and both models sit at 90.9 on GPQA Diamond ^[19].