DeepSeek has shipped V4.1-Flash, and the spec sheet reads like two different models argued and neither won. It is a 552-billion-parameter mixture-of-experts system with a million-token context window and native vision, and it serves cached input tokens at three tenths of a cent per million during off-peak hours.
For comparison, the frontier labs charge somewhere between $0.40 and $0.50 per million cached input tokens. That is a difference of more than two orders of magnitude, and the weights are open under an MIT license.
The Architecture Is the Interesting Part
V4.1-Flash is built on what DeepSeek calls a causal encoder-decoder design, splitting 40 layers into a 20-layer encoder and a 20-layer decoder. The model activates 8 billion parameters while processing input and 16 billion while generating output, an asymmetry the previous Flash model did not have. V4-Flash activated a flat 13 billion for both phases.
That split is a direct optimization for how these models actually get used. Prefill, reading a long prompt, is compute-bound and embarrassingly parallel. Decode, writing the answer one token at a time, is memory-bandwidth-bound and serial. Giving decode more active parameters and prefill fewer means you spend capacity where it changes output quality and save it where it does not.
The Cache Numbers Are the Real Story
Buried in the specs is the figure that explains the pricing: the KV cache footprint is roughly 890 bytes per token, about a quarter of what V4-Flash required. There are also 196 billion parameters sitting in sparsely accessed memory modules rather than in the dense path.
Serving costs for long-context inference are dominated by KV cache memory. Cut that by 75% and you can hold four times as many concurrent conversations on the same hardware. The $0.003 cached input price is not a loss-leader stunt, it is what the architecture permits.
What It Costs
DeepSeek has kept its unusual time-of-day pricing. Off-peak rates:
| Token type | Off-peak price per 1M |
|---|---|
| Cached input | $0.003 |
| Uncached input | $0.15 |
| Output | $0.60 |
Peak rates are exactly double across the board. The peak windows run Monday to Friday, 01:00 to 04:00 and 06:00 to 10:00 UTC, which tracks Chinese business hours. Anyone running batch workloads from North America or Europe will find the cheap window lines up neatly with their own off-hours.
Output at $0.60 per million is the number that will reshape budgets. Agentic workflows, the ones that loop dozens of times and generate far more than they consume, have been economically painful at frontier pricing. At these rates they become almost free to iterate on.
Benchmarks: Close, With Caveats
DeepSeek reports a 74.2 on DeepSWE v1.1, marginally ahead of the 74.0 it reports for Claude Opus 5. On Terminal-Bench, Anthropic’s model keeps the lead.
Treat a 0.2-point gap on a self-reported agentic benchmark as a tie, because that is what it is. But a tie is itself the headline. An open-weight model that trades evenly with a frontier proprietary system on agentic coding, at roughly a hundredth of the input cost, changes the calculus for anyone who was paying frontier rates out of necessity rather than preference.
The honest caveat is that benchmark parity and production parity are different animals. Frontier models tend to hold up better on the long tail of weird inputs, tool-use edge cases and instruction-following under pressure. Those gaps do not show up in a score table. They show up three weeks into a deployment.
The Deprecation Nobody Asked For
DeepSeek retired the legacy V4-Flash identifiers, which now silently route to V4.1-Flash. V4 Pro has likewise been pulled in favour of the new model.
Silent model substitution behind a stable API name is a real operational hazard. If you have prompts tuned against V4-Flash behaviour, evaluation suites calibrated to its outputs, or downstream parsers expecting a particular response shape, your pipeline changed underneath you without a deploy. Teams that pin model versions explicitly are fine. Teams that did not should be re-running their evals this week.
Running It Yourself
The weights are on Hugging Face under an MIT license, which permits commercial use without restriction. That is the most permissive license in the open-weight tier and a meaningful differentiator against models that ship with usage caveats.
Self-hosting is a different matter. A 552-billion-parameter MoE, even with low activation counts, needs serious memory to hold the full weight set. This is multi-GPU-node territory, not something that fits on a workstation. For most organizations the practical path is the API, with self-hosting reserved for those with hard data-residency requirements and hardware to match.
Final Verdict
V4.1-Flash is the clearest evidence yet that the cost floor for capable AI inference is still falling fast, and that the fall is being led by open weights rather than the labs with the biggest training budgets.
The architectural work here is genuinely good. The asymmetric activation and the KV cache reduction are not marketing, they are the reason a million-token context model can be served at these prices. Anyone building agentic systems, document processing at volume, or anything that burns output tokens by the million should be testing this against whatever they currently pay for.
Where it falls short of a clean win: the benchmark lead is inside the noise, Terminal-Bench still favours the incumbents, and the silent deprecation of older identifiers is a reminder that cheap comes with operational friction. Price it in. Then run your own evaluation, because at $0.003 per million cached tokens, the test is nearly free.
