DeepSeek releases V4.1 Flash - a million-token context, images, and prices dozens of times lower than Anthropic's

Every DeepSeek launch follows the same script. The company shows charts where its model beats OpenAI and Anthropic, and then the whole internet spends a few days checking how much of it is true. Today, September 10, DeepSeek released V4.1 Flash - the smallest model of the new V4.1 family and, as the company's researcher Zizheng Pan wrote on X, "our first flagship model with native multimodal support." The weights are on Hugging Face under the MIT license, so anyone can run the model themselves. DeepSeek's post on X had 3.4 million views after twelve hours.

The numbers impress mostly through contrast. The model has 552 billion parameters in a mixture-of-experts architecture, but only 8 billion parameters are active per token when reading the input and 16 billion when generating the answer - DeepSeek calls it an asymmetric Causal Encoder-Decoder design. On top of that come 196 billion parameters of Engram memory, which can sit on an ordinary SSD. The context window is a million tokens, or several thick books at once, and a response can run to 384,000 tokens. The KV cache takes up 890 bytes per token, a quarter of the previous Flash, and that is a large part of why long conversations are cheap. Hacker News users measure 250-400 tokens per second from the API.

Off-peak pricing: 0.02 RMB per million input tokens read from the cache, 1 RMB for regular input and 4 RMB for output - about $0.15 and $0.60. During peak hours, Monday through Friday from 9:00 to 12:00 and 14:00 to 18:00 Beijing time (01:00-04:00 and 06:00-10:00 UTC), the rates double. For comparison, Claude Fable 5 costs $10 per million tokens for input and $50 per million for output, so off-peak DeepSeek is 66 and 83 times cheaper, respectively. It shows most clearly on long agentic tasks. User k9294 on Hacker News priced out the token usage from one coding session (447 turns, 36.5 million tokens read from the cache) under both models' price lists: $55 at GPT-6 Astra rates, 36-73 cents at V4.1 Flash rates. Another user, mmastrac, pushed 2.1 billion tokens through the preview version in two days and paid $22.

Artificial Analysis calculated in August that a single task from its test set cost 3 cents on the previous V4 Flash and $3.15 on Fable 5, more than a hundred times as much. But that calculation was still based on DeepSeek's old, lower prices. The first independent tests of the new model appeared on X within hours. Paweł Huryn gave 23 models two repositories with 105 hidden bugs to fix. Selected results: GPT-5.6 Luna (max) fixed 33 for $1.80, Claude Opus 5 (max) 27 for $51.33, Grok 4.6 (xhigh) 27 for $16.96, DeepSeek V4.1 Flash (max) 24 for $1.08. The aforementioned mmastrac summed up after two days of work: "slightly north of Opus 4.8, not quite Opus 5, not Fable."

So what about "beating OpenAI and Anthropic"? In DeepSeek's table on Hugging Face, the new model beats Claude Opus 5 in DeepSWE (74.2 to 74.0), Terminal-Bench 2.1 and AutomationBench, but loses in GPQA Diamond (90.9 to 93.4), Terminal-Bench 3.0 (30.0 to 43.3) and Humanity's Last Exam without tools (36.8 to 56.3). In the latest version of the independent Intelligence Index, released September 7, the previous V4 Flash scores 35 and August's V4 Pro scores 36, while Claude Opus 5 scores 51, Fable 5 scores 50 and GPT-5.6 Sol scores 47. Artificial Analysis has not measured V4.1 Flash yet. It is more honest to talk about the best price-to-performance ratio than about a better model. Though at this price gap there is still plenty to talk about.

There is a flip side. The previous V4 Flash had 284 billion parameters; the new one has almost twice that. "It's not really flash anymore," wrote Hacker News user revolvingthrow, and another, petu, estimated that usable speed at home needs about 384 GB of memory, meaning three Nvidia Spark machines or four RTX Pro 6000 cards. Others replied that flash means speed, not size, and this model is twice as fast as its predecessor. Both sides are right - the model simply no longer fits on a laptop.

An interesting detail from the notice to API customers: from September 14 (04:00 UTC) all requests to V4 Pro will be routed to V4.1 Flash and billed at Flash rates. DeepSeek writes that tests by multiple parties gave the new, smaller model the edge over the old flagship in performance, cost, speed and total runtime. V4.1 Pro is due later, with no date given. The model is already supported by Tencent (WorkBuddy, CodeBuddy) and OpenCode.

It is worth keeping the calendar in mind. In May, DeepSeek announced that the 75-percent price cut on V4 Pro would be permanent. In mid-August it raised prices by 50 to 1,100 percent, introduced peak/off-peak tariffs and priced V4 Pro at roughly three times the Flash rate. According to reports, the company is raising about $8 billion at a $74 billion valuation, preparing a listing on Shanghai's STAR Market, building data centers and designing its own chips. Today's cut is 32 percent on input, 57 percent on the cache and 9 percent on output, while MiniMax shares fell 9 percent in Hong Kong, Z.ai 10 percent and Alibaba more than 3 percent. "Cheapest" is therefore a relative term - and depends on the time of day.

Mixture of experts (MoE) - an architecture in which the model consists of many specialized sub-networks ("experts") and activates only a few of them for each token. This lets the model be enormous while computing as fast and cheaply as a model dozens of times smaller. KV cache (key-value cache) - the place where the model keeps the already-processed content of a conversation so it does not have to recompute it for every new word; the smaller it is per token, the cheaper long sessions are.