DeepSeek released V4.1 Flash on September 10, and the timing says something about how the company wants to compete. Instead of chasing a single benchmark crown, it built a model around a specific bet: that the real cost of running AI agents comes from rereading long contexts over and over, not from the first prompt a user types. The result is a 552-billion-parameter mixture-of-experts model with native vision, a context window of one million tokens, and open weights published under the MIT license on Hugging Face, free for commercial use.

The new pricing math

The headline number is the price. During off-peak hours, which DeepSeek defines as everything outside 01:00–04:00 UTC and 06:00–10:00 UTC on weekdays, V4.1 Flash costs $0.003 per million input tokens on a cache hit, $0.15 per million on a cache miss, and $0.60 per million output tokens. At peak hours, all three figures double.

That distinction sounds technical, but it changes how developers should think about agent costs. A coding assistant or research agent that keeps reusing the same repository, system instructions, or conversation history reads far more cached context than fresh context. DeepSeek says cache-hit charges already make up a large share of typical agent spending, which is why it built pricing around that pattern instead of around raw token counts.

Compared to other frontier APIs, the gap is wide. OpenAI charges $4 per million regular input tokens for GPT-5.6 Sol, with cached input at $0.40 and output at $20. Anthropic prices Claude Opus 5 at $5 for standard input, $0.50 for cache hits, and $25 for output. Moonshot AI’s Kimi K3 sits at $3 for cache-miss input and $15 for output. Even at peak hours, V4.1 Flash tops out at $0.30 for uncached input and $1.20 for output, a fraction of any of those rates. For an agent that reuses a 500,000-token prefix across 100 requests, that pricing difference alone can mean a bill of roughly $0.15 on V4.1 Flash versus $20 or more on the frontier competitors, based on their published cache rates.

Under the hood: a different kind of Flash

DeepSeek’s “Flash” naming used to mean smaller. That is no longer accurate here. The previous V4 Flash ran on a 284-billion-parameter backbone with 13 billion active parameters. V4.1 Flash grows the backbone to 552 billion, split into a 20-layer causal encoder and a 20-layer decoder, a design DeepSeek calls Causal Encoder-Decoder. It activates 8 billion parameters while reading input and 16 billion while generating output, so the earlier stage got lighter while the generation stage got heavier.

The efficiency claim lives in the cache, not the parameter count. Combined with a technique DeepSeek calls Compressed Sparse Attention 2, hierarchical sparse indexing, and FP4 KV caching, the company says the global key-value cache drops to about 890 bytes per token, roughly a quarter the size of V4 Flash’s, while persistent storage needs fall to about an eighth. DeepSeek also reports that prefill computation for long sequences drops by close to half, while decode compute per token grows only around 25% as context scales 256-fold, from 4,000 tokens to a million.

DeepSeek is upfront about the tradeoffs. The company’s technical report acknowledges that the new architecture introduces untested edge cases, including possible errors in sparse attention selection and in the approximate memory reconstruction used when short-lived cache states expire. DeepSeek says it has not observed systematic problems in internal testing so far, but plans further stress testing before treating the design as fully proven.

Where the benchmarks land

DeepSeek reports a 74.2 score for V4.1 Flash on DeepSWE v1.1, ahead of the 74.0 it lists for Claude Opus 5 and 73.0 for GPT-5.6 Sol, along with 88.1 on CyberGym and 54.8 on AutomationBench. Those numbers come from DeepSeek’s internal testing, run at its maximum reasoning-effort setting, and they do not paint a picture of total dominance. On Terminal-Bench 3.0 and 4.0, Claude Opus 5 leads V4.1 Flash by wide margins, and GPT-5.6 Sol comes out ahead on GPQA Diamond and SEC-Bench Pro in the same comparison table.

Reasoning effort turns out to be a separate cost lever. DeepSeek says raising effort from 25 to 100 lifts DeepSWE v1.1 from 66.0% to 74.2%, but uses roughly 2.5 times as many output tokens to get there. Effort levels between 60 and 80 recover most of the accuracy gain at less than half that token cost, which means the model’s best setting for a given task may not be the one behind its published leaderboard scores.

The pushback from developers

Not every reaction has been positive. On Hacker News, developers questioned whether a 552-billion-parameter model can still be called “Flash,” pointing out that self-hosting the open weights now takes considerably more hardware than before, even if server-side efficiency improves. A separate complaint targeted DeepSeek’s decision to automatically route existing V4 Flash API calls to V4.1 Flash, and to do the same for V4 Pro calls starting September 14 until a V4.1 Pro model appears. Developers argued that swapping the model behind a stable API identifier can break tuned prompts and pass regression testing that was never designed to detect a silent model change.

A listing in the background

The release lands as DeepSeek’s business plans widen. Reuters reported this week that the company has engaged CITIC Securities to prepare for a possible listing on Shanghai’s STAR Market, with a fundraising round that could value DeepSeek at close to 500 billion yuan, around $75 billion. Seen against that backdrop, V4.1 Flash reads less like an attempt to win every benchmark and more like a demonstration that DeepSeek can serve capable models at a fraction of the going rate, at a moment when that argument carries financial weight.

For developers deciding whether to switch, the benchmarks are only part of the picture. The bigger question is whether a given workload is repetitive and cache-heavy enough to take advantage of V4.1 Flash’s pricing, and whether the savings outweigh the cost of testing a model whose edge behavior is still being mapped.