DeepSeek's V4.1-Flash costs $0.15 per million tokens for four-fifths of the week
The MIT-licensed release ties Opus 5 on DeepSWE, trails it badly on the hardest terminal benchmark, and the checkpoint is 38 percent bigger than the number in the headlines.
DeepSeek's new model has two prices, and the one in the headlines applies for about a fifth of the week.
V4.1-Flash arrived on 10 September 2026 under an MIT licence, with a million-token context and native vision. The rate card says $0.30 per million input tokens and $1.20 per million output. Then read the footnote. Peak hours are two weekday blocks in the middle of the Chinese working day, and everything outside them costs half. Add the blocks up and peak covers 35 hours out of 168.
So for anyone outside those office hours — which is to say most of the people reading this — the real rate is $0.15 in and $0.60 out. Cached input off-peak is priced at $0.003 a million, which 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 far larger model, at roughly a quarter of the cost per token.
The checkpoint is bigger than the number everyone repeated
We spent the afternoon in the model card and the repository rather than the launch thread. Start with the parameter count, because everyone is repeating a backbone figure of 552B as though it were the model.
The card also lists Engram conditional memory at 196B parameters, "sparsely accessed via token-based lookup" — a 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. The thing you download is about 38 percent larger than the headline, and roughly a quarter of it is memory rather than computation.
That is the trick, and it is 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.
The benchmark table is more honest than its fans
DeepSeek publishes that chart itself. On DeepSWE the new model ties Opus 5 at just over 74 percent of issues resolved, and on the older Terminal-Bench it leads the field outright. On the harder, newer one, Opus is twenty points clear. The agentic wins are real. The reasoning gap is real too, and only one of them is in the posts going around.
There is a subtler cost as well. Artificial Analysis measured Flash at 89,000 tokens per Intelligence Index task, the most verbose model it has on record, more than either American lab's flagship. A model that costs a quarter as much per token and talks three times as long is not a quarter as cheap. It still 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.
The price card is the release
The architecture is a cache paper and the benchmark table is a strong second tier, and neither would matter much alone. What matters is that a model that ties the leading coding model on DeepSWE now costs sixty cents per million output tokens for five days and thirteen hours out of every seven.
So what does that do to a coding product sold at a flat monthly price? Its operator has to decide whether to keep paying frontier rates for a difference most 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 this model displaces Opus on hard, long-horizon work this year. The gap on the harder benchmark is too wide.
Independent scoring lands in the same place. OpenDesign Arena put Flash within two points of GPT-6 Astra on real design tasks while charging a fraction of a cent per finished design against Astra's dollar and change, and eleven of the thirteen models it ran scored lower 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", which is a reasonable choice and also a choice. Somebody should rerun the hard benchmark 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.
