Published 10 September 2026
DeepSeek released DeepSeek-V4.1-Flash on 10 September 2026 and cut its API price to $0.15 per million input tokens off-peak, with output at $0.60. The model has a 552 billion parameter backbone but switches on only 8 billion of them for each token it reads. The weights are on Hugging Face under the MIT license, and the context window holds one million tokens.
What is DeepSeek-V4.1-Flash and why is it so cheap to run?
DeepSeek-V4.1-Flash is a mixture-of-experts model, which means it switches on only part of itself for each word. It reads images as well as text, and writes text back. DeepSeek trained it from scratch on 45 trillion tokens of mixed text and images.
The cost story is about the KV cache, the working memory a model keeps for every token in your prompt. That memory is what makes long agent runs expensive, because it grows with the conversation and has to stay in GPU memory. DeepSeek says V4.1-Flash holds 890 bytes per token, about a quarter of what DeepSeek-V4-Flash needed and, by its own chart, roughly 437 times less than DeepSeek-V1.
Two design choices do that work. A causal encoder-decoder layout splits the 40 layers into a 20-layer encoder and a 20-layer decoder, so the decoder builds its cache once from the encoder output instead of one cache per layer. That is why the model activates 8 billion parameters while reading a prompt and 16 billion while writing. On top of that, DeepSeek stores the cache in FP4, a four-bit number format, and shares attention indices between layers. You also get a reasoning effort dial from 1 to 100, so you can trade accuracy against cost per request rather than switching models.
DeepSeek-V4.1-Flash benchmarks against Claude Opus-5.0
DeepSeek published a full comparison table. Read it as two different stories: on the coding and agent tests that the industry has been quoting all year, V4.1-Flash matches a frontier model. On the newest and hardest tests, it does not.
| What is measured | DeepSeek-V4.1-Flash | Claude Opus-5.0 |
|---|---|---|
| DeepSWE v1.1 (fixing real bugs in real repositories) | 74.2 | 74.0 |
| Terminal-Bench 2.1 (getting work done in a terminal) | 90.6 | 89.1 |
| Terminal-Bench 4.0 (the harder terminal set) | 31.2 | 51.8 |
| Agent’s Last Exam (long multi-step agent tasks) | 31.8 | 28.6 |
| GPQA Diamond (PhD-level science questions) | 90.9 | 93.4 |
| Humanity’s Last Exam (hardest expert questions) | 36.8 | 56.3 |
These are DeepSeek’s own numbers. It ran the coding tests through its own DeepSeek Harness in Minimal mode with a one million token context and the reasoning dial at 100, and it names the harness for every row. What the model card does not say is where the Opus-5.0 and GPT-5.6 Sol figures came from, so treat the rival columns as collected rather than rerun.
DeepSeek did publish one thing most labs leave out: how much the harness changes the answer. The same model scores 74.2 on DeepSWE v1.1 under mini-SWE, 72.6 under its own harness, 69.8 under Claude Code and 65.6 under Codex. The headline number is the best of eight scaffolds, and the spread is nearly nine points. If you run your agents on Codex, plan for the bottom of that range.
What DeepSeek is not saying about DeepSeek-V4.1-Flash
The comparison table has no column for GPT-6 Astra, the model OpenAI shipped a week earlier and re-announced for business use on 9 September. OpenAI reports 57.9 on Terminal-Bench 4.0 for Astra. DeepSeek reports 31.2 for V4.1-Flash on the same test and compares itself to GPT-5.6 Sol instead. That is the missing row, and it is the one that matters if you are choosing a coding agent today.
Two practical gaps are worth knowing before you download anything. The release ships no chat template in the usual Jinja format, so you either run DeepSeek’s Python reference script or adopt its new Rust toolkit to build prompts correctly. And DeepSeek publishes no hardware requirement for self-hosting a 552 billion parameter checkpoint, nor any tokens-per-second figure, which is a strange omission for a release whose whole argument is efficiency.
Price, licence and what happens to DeepSeek V4 Pro
- Licence: MIT, weights published on Hugging Face with a technical report
- Off-peak price: $0.15 per million input tokens, $0.60 per million output
- Peak price: $0.30 and $1.20, during 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays
- Cached input: $0.003 off-peak, $0.006 peak, per million tokens
- Context: one million tokens in, up to 384,000 out, images and text supported
- API model name: deepseek-flash, with a concurrency limit of 2,500 requests
The older DeepSeek-V4-Flash and DeepSeek-V4-Flash-Vision-Exp are retired, and their names now point at V4.1-Flash. From 12:00 Beijing time on 14 September 2026, calls to deepseek-v4-pro also route to V4.1-Flash and get billed at the Flash price, which is roughly a quarter of what V4 Pro cost. DeepSeek says a V4.1 Pro will follow but gives no date. If you built anything on V4 Pro, that is a model swap you did not choose, arriving in four days.
What this means
This is worth testing now if you run coding or agent work at volume and your bill is driven by long prompts. Think of a team running code review agents over a large repository, where every request re-reads thousands of lines of context. At $0.15 per million input tokens off-peak, with cached input at $0.003, that workload gets cheaper by a factor most teams would notice in a week. The MIT licence means you can also run it on your own hardware with no negotiation, which matters for anyone who cannot send source code to a Chinese API.
It is not the model to put behind a research assistant or anything that has to answer hard factual questions. A score of 36.8 on Humanity’s Last Exam against 56.3 for Opus-5.0 is not a rounding error, and the 20-point gap on Terminal-Bench 4.0 says the same thing in a different way. The honest read is that DeepSeek has built a very cheap model that is excellent at the agent tasks of six months ago and mid-table on the ones being set now. Test it on your own repository, with your own harness, before you move any traffic. The first thing we would measure is cost per completed task rather than price per token, because the scaffold spread DeepSeek published suggests your result will depend as much on your agent framework as on the model.
For more information, visit the official announcement of DeepSeek-V4.1-Flash in the DeepSeek documentation.
