Three-tenths of a cent buys a million tokens of cached input on DeepSeek V4.1 Flash off-peak. The same million tokens cost $0.15 as a cache miss, and $0.60 as output. Within a single pricing tier, a cache hit is fifty times cheaper than a miss — and that ratio, not the parameter count, is the most important number released this week.
Last issue tracked GLM-5.3's approach to parity and the argument that open weights were beginning to reprice inference. This week DeepSeek V4.1 Flash, MiniMax M3 and Kimi K2.8 stopped arguing and moved the floor again.
DeepSeek shipped V4.1 Flash on September 10 under an MIT license. It is a 552-billion-parameter mixture-of-experts model with a 1M-token context window and native multimodal input, and its weights are on Hugging Face.
It is a 552-billion-parameter mixture-of-experts model with a 1M-token context window and native multimodal input, and its weights are on Hugging Face.
The architecture is the interesting part. DeepSeek calls it a Causal Encoder-Decoder: 40 transformer layers split into 20 encoder and 20 decoder layers, activating roughly 8B parameters during prefill and 16B during decode.
That asymmetry is a commercial design, not just a technical one. You get a heavyweight backbone at a lightweight compute cost per token, which is precisely what an agent that re-reads a repository on every loop needs.
The KV cache shrinks to one quarter the HBM and one eighth the SSD footprint of the prior generation, at a reported 890 bytes per token. Read that figure precisely: it describes the global KV cache, not total VRAM or a minimum server configuration.
The MIT license is the part that changes procurement. Prior DeepSeek open-weight releases carried commercial restrictions; this one has no revenue thresholds and no enterprise agreements. For regulated and privacy-sensitive workloads that previously had no viable open-weight option, that is a real unlock.
The hardware floor is also real. The checkpoint weighs 510 GB across 48 files, with a minimum self-hosting build of 8x H100s at FP8; FP4 quantization drops it to roughly 385 GiB, and a single B200 covers the native FP4+FP8 build. This is not a workstation model.
Pricing runs on a peak/off-peak schedule. Off-peak is $0.003 per million cached input tokens, $0.15 for misses, $0.60 for output. Peak doubles all three, and peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays.
At off-peak rates, output costs roughly 40 times less than Claude Opus 5. On a fixed illustrative invoice of one million input tokens and 100,000 output tokens, moving from zero cache hits to 90% cuts the bill from $0.42 to $0.1554 — a 63% reduction.
That illustration is the whole strategic point. DeepSeek is charging developers to learn cache discipline, because the 50-fold spread between hit and miss makes application architecture visible in the invoice.
The agent market is becoming a memory market. Input list price matters less than whether a long-running workflow can reliably reuse its prefix, and how much context it has to rebuild after every tool call, retry and verification step.
The vendor benchmarks look strong. CyberGym: 88.1 against GPT-5.6 Sol's 84.5. Codeforces rating of 3471, up from V4-Pro's 3348. Terminal-Bench 3.0 at 30.0, against V4-Pro's 11.8.
Together AI, hosting the model, put it more bluntly: V4.1 Flash beats GPT-5.6 Sol on agentic benchmarks at one third the cost per task. That claim deserves scrutiny.
The same thread did the arithmetic: $100 divided by $0.27 is 370, so the "3x tasks" framing is a price ratio repackaged. The $0.27 figure is an Intelligence Index item, not a Terminal-Bench episode, and cost per solved agent trajectory remains unpublished.
That is the gap in this entire week's story. Cache-hit ratios are list prices; they are