DeepSeek V4.1 Flash rebuilds the cache for long-context agents, and quietly becomes the API default on September 14

DeepSeek released V4.1 Flash this week as open weights under an MIT license, and the interesting part is not the benchmark table. It is a native multimodal model with a one million token context window, but the design choices underneath are aimed squarely at one workload: long-running agents that re-read a large context on every step. DeepSeek rebuilt the way the model stores that context, and the result is a model that gets cheaper exactly where agent pipelines are most expensive to run. On September 14 it also becomes the default that answers a large share of DeepSeek API traffic, whether or not you asked for it.

The cache is the point

Frontier serving cost is dominated less by raw compute than by the key-value cache, the running memory of everything the model has already read in a session. For an agent that loops over a big system prompt, tool definitions, and prior steps, that cache grows every turn and has to be held in expensive accelerator memory. V4.1 Flash attacks it directly. DeepSeek reports a global KV cache of about 890 bytes per token, roughly a quarter of the previous V4-Flash, achieved through FP4 quantization of the cache plus a causal encoder-decoder layout that reuses attention across layers rather than recomputing it. The mixture-of-experts backbone is 552 billion parameters, but only about 8 billion are active while the model reads input and 16 billion while it generates, so the cost of feeding it a long context stays low.

None of that shows up in a leaderboard screenshot, and that is precisely why it matters. The bottleneck for input-heavy agent work has been serving economics, not whether the model can reason. A model that holds a million-token context on a fraction of the memory changes what a small team can afford to leave running.

The API default changes under you on September 14

The pricing move is the concrete consequence. DeepSeek’s published rates put Flash output at $0.60 per million tokens off-peak and $1.20 at peak, with cache-hit input as low as $0.003 per million. For comparison, Claude Opus 5 lists $25 per million output tokens, so DeepSeek is undercutting the closed frontier by more than an order of magnitude for teams whose workloads tolerate a non-Western model. The peak window is a narrow band of weekday UTC morning hours; everything else bills at the cheaper off-peak rate.

The part worth flagging is operational. DeepSeek’s docs state that from September 14, requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at the Flash price until a future V4.1 Pro ships. Anyone calling the Pro endpoint in production gets a different model with a different behavior profile without changing a line of code. That is a config-level event masquerading as a pricing note, and it is the kind of silent substitution that a regression test catches and a release note does not.

Multimodal and open, on the frontier’s terms

V4.1 Flash reads images and text natively rather than bolting vision on afterward, and DeepSeek says it trained on 45 trillion tokens at a 7-to-1 text-to-multimodal ratio. The weights ship openly with vLLM, SGLang, and Transformers serving paths, so a team that does not want to route sensitive data through a hosted API can run it in-house, which is the option a closed model never offers.

On capability, the honest read is mixed. DeepSeek’s own tables show V4.1 Flash landing ahead of or close to GPT-5.6 Sol and Claude Opus 5 on several agentic and coding measures, including a narrow edge on DeepSWE v1.1, and trailing them on newer terminal-agent and knowledge-heavy tests. Those are vendor-reported numbers, and independent evaluations have not yet reproduced them. Read the release as a cost and serving story with competitive-but-not-leading quality, not as a claim to the top of the frontier.

What it means for developers

Three groups feel this immediately. Teams already on the DeepSeek API should decide before September 14 whether the Pro-to-Flash swap is acceptable, and pin an explicit model id if their pipeline depends on Pro’s behavior. Teams building input-heavy agents get a genuinely cheaper substrate for the workload that has been hardest to run at scale, provided the quality holds on their own evals rather than DeepSeek’s. And teams with data-residency or self-hosting constraints get an open-weight multimodal model with a million-token window they can run on their own hardware, which reframes the build-versus-buy math that closed frontier pricing had mostly settled.

What to watch

The signals over the next quarter are whether independent benchmarks reproduce the agentic-coding results outside DeepSeek’s tables, whether the 890-byte cache figure holds up under real long-session agent traffic rather than clean measurements, and how V4.1 Pro is positioned when it lands, given that Flash is now carrying the Pro endpoint’s load. The broader pattern is that the cheapest serving economics for long-context agents are increasingly coming from open weights, which pressures the closed labs on exactly the workload they most want to own. Stackmaven’s follow-up will revisit those questions around December 10.

Sources cited

1. DeepSeek-V4.1-Flash model card (Hugging Face, DeepSeek) huggingface.co
2. Models and pricing (DeepSeek API docs) api-docs.deepseek.com
3. DeepSeek released DeepSeek-V4.1-Flash with 1M context, FP4 KV cache, and cross-layer attention reuse (MarkTechPost) www.marktechpost.com
4. DeepSeek V4.1 Flash matches GPT-5.6 Sol and Claude Opus 5 on some benchmarks at far lower pricing (OfficeChai) officechai.com