DeepSeek releases V4.1-Flash

DeepSeek shipped V4.1-Flash on September 10. Open weights on Hugging Face, MIT license. It is the smallest model in a new architecture family, and DeepSeek's claim is that it beats the flagship V4-Pro on performance, cost, speed and total runtime. 48K posts on X in a day.

I read the model card, the API docs and the coverage. Three things: where it actually improved, what it costs, and one decision DeepSeek reversed within 24 hours.

552B parameters, 8B active per token

The architecture is a Causal Encoder-Decoder: 40 layers, 20 encoder and 20 decoder. 552B total parameters, 8B active per token during prefill, 16B during decode. The previous V4-Flash had 284B total. Parameters doubled, active compute did not. That is the first source of the low price.

The second is the cache. The KV cache is stored in 4-bit floating point at 890 bytes per token, about a quarter of V4-Flash. On long-context work that number sets the cost of a call.

Context is 1M tokens, max output 384K. Image input is now built into the model instead of an experimental variant.

Benchmarks: ahead of Opus 5 on two, behind on one

From the official model card at max reasoning effort:

DeepSWE v1.1 (tasks resolved): 74.2%. V4-Pro 62.7%, Claude Opus 5 74.0%, GPT-5.6 Sol 73.0%.
Terminal-Bench 2.1: 90.6. V4-Pro 87.9, Opus 5 89.1, Sol 88.8.
GPQA Diamond: 90.9. V4-Pro 92.4, Opus 5 93.4, Sol 94.1.
Codeforces rating: 3471. V4-Pro 3348.

The shape is clear. On coding and terminal work it passes its own flagship and matches or edges the two closed frontier models. On the knowledge benchmark it is 1.5 points below V4-Pro. Same shape as GPT-6 Astra last week: this generation is growing toward doing work, not knowing more.

All of this is self-reported. No third-party evaluation yet.

Price: $0.60 per million output tokens, 30% of V4-Pro

On the API the model is now called deepseek-flash. Off-peak: $0.003 per million input on cache hit, $0.15 on cache miss, $0.60 per million output. Peak hours double that. Peak is 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday.

V4-Pro off-peak: $0.66 input on cache miss, $1.98 output. So output drops to 30%.

A model that scores higher than the flagship, sold at 30% of the flagship's price. That is the real pricing move.

The decision reversed in 24 hours

On September 9 DeepSeek wrote on its site that from September 14 all deepseek-v4-pro requests would be routed to V4.1-Flash and billed at Flash rates until V4.1-Pro ships. Several outlets reported it that way on the 10th.

The API changelog on the 10th says something else: "We have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged."

From forced migration to keeping the old model in one day, with no explanation. The most likely reading is that customers refused to have their model swapped silently, even for a higher-scoring one. That holds for every company selling models: customers buy predictability, not scores.

Same day, other news

Anthropic's threat report on the 10th named seven Chinese labs for collecting Claude outputs through fraudulent accounts to train on. DeepSeek is on the list, with a figure of more than 12.1 million exchanges over 14 days in July. DeepSeek has not responded. I am relaying this, not judging it.

A test that does not favor me

Our gateway lists no DeepSeek model at all. Calling deepseek-flash or deepseek-v4-pro returns "no available channel." The cheapest tier of models in China, and we have none of them wired up.

Caveats

All benchmark numbers are DeepSeek's own. "Multiple parties tested" names no parties. The routing reversal has a single source, the changelog, with no official statement.

Disclosure: I run an LLM gateway. The cheaper DeepSeek gets to call directly, the worse for me.

If a model with higher scores at 30% of the price silently replaced the flagship you use tomorrow, would you accept it?