DeepSeek's V4.1-Flash costs $0.15 per million tokens for four-fifths of the week

DeepSeek released V4.1-Flash on September 10 at $0.30 per million input tokens and $1.20 per million output, MIT-licensed, with a million-token context and native vision. Then read the footnote on the price page. Peak hours run 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, and everything outside them costs half. That is 35 hours out of 168, so the advertised price applies to 20.8% of the week and the real rate for most of the world most of the time is $0.15 in and $0.60 out. Cached input off-peak is $0.003 per million.

Which of those is the real price? The off-peak one, for anybody outside Chinese office hours — which is to say most of the people reading this. It is less a price than a rounding error with a decimal point in it.

Artificial Analysis put the model at 40 on its Intelligence Index, ahead of DeepSeek's own V4 Pro, a 1.6-trillion-parameter model, at roughly a quarter of the cost per token. Keep their cost-per-task figure.

We spent the afternoon in the model card and the repository rather than the launch thread, and two things there are worth your time.

Start with the parameter count. Everyone is repeating 552B. But 552B is only the backbone. The card also lists Engram conditional memory at 196B parameters, "sparsely accessed via token-based lookup" — a lookup table the model reads instead of computing. Add the vision tower and the rest and the safetensors index on Hugging Face totals 763,205,315,794 parameters across 510 GB of weights. So the thing you download is about 38% larger than the number in every headline, and roughly a quarter of it is memory rather than computation. And that is the trick, a good one: the activated path is 8B parameters while the model reads your prompt and 16B while it writes, which is why input-heavy agent loops get cheap.

Then the benchmark table, which is more honest than its fans.

DeepSeek publishes that chart itself. On DeepSWE the model resolves 74.2% against Opus 5's 74.0%, which is a tie, and on Terminal-Bench 2.1 it leads the field at 90.6. But on Terminal-Bench 4.0 — the harder, newer one — Opus 5 scores 51.8 and V4.1-Flash 31.2, a gap of 20.6 points. On Humanity's Last Exam without tools it is 36.8 against Opus 5's 56.3. The agentic wins are real; the reasoning gap is real too, and only one of them is in the posts going around.

And there is a subtler cost. Artificial Analysis measured V4.1-Flash at 89,000 tokens per Intelligence Index task, the most verbose model they have on record, 62% more than DeepSeek's own V4 Pro and more than Fable 5.1 or Opus 5. A model that costs a quarter as much per token and talks three times as long is not a quarter as cheap. Even so, it comes out at $0.27 a task against $2.01 for GLM-5.3, so the arithmetic survives — but if you are budgeting agent runs, budget the tokens, not the rate card.

Our read: the price card is the release. Its architecture is a KV-cache paper (890 bytes per token, roughly a quarter of V4-Flash's) and the benchmark table is a strong second tier, and neither would matter much on its own. What matters is that a model tying Opus 5 on DeepSWE now costs $0.60 per million output tokens for five days and thirteen hours out of every seven. Somebody selling a coding agent at a flat monthly price has to decide whether to keep paying frontier rates for a difference their users cannot see on most tickets. We would expect at least two established coding products to add a DeepSeek route by the end of October, and we would take the other side of any claim that V4.1-Flash displaces Opus on hard, long-horizon work this year — the Terminal-Bench 4.0 gap is too wide.

Independent scoring lands in the same place. OpenDesign ran thirteen models through the same design tasks and gave Astra 82.7 and V4.1-Flash 81.2, at $1.61 and $0.023 per finished design respectively; eleven of the thirteen scored lower than DeepSeek and cost more.

One caveat we would like closed. Every code-agent number in the comparison table comes from DeepSeek's own harness in "Minimal mode" with a 1M-token window, which is a reasonable choice and also a choice. Somebody should rerun Terminal-Bench 4.0 on a neutral scaffold. Until then, the cheapest honest summary of this release is the one on the pricing page, not the one in the benchmark table.