DeepSeek V4.1 Flash: Pricing, Specs, V4 Pro Routing, and How to Access It (2026)

DeepSeek replaced its whole API lineup with one model this week. Here’s what V4.1 Flash costs, what’s inside it, what happens to your V4 Pro calls on Sunday, and how to run it.

DeepSeek V4.1 Flash went live on the official API on September 10, 2026. The model name is deepseek-flash, the price is $0.15 per million input tokens and $0.60 per million output off-peak, double that at peak, and cache hits are $0.003. It replaces V4 Flash outright: the old deepseek-v4-flash and deepseek-v4-flash-vision-exp names still resolve, but the models behind them are retired. From noon Beijing time on September 14, it also takes over V4 Pro. Every request to deepseek-v4-pro will be served by V4.1 Flash and billed at V4.1 Flash rates until a V4.1 Pro exists, and DeepSeek hasn’t given a date for that.

DeepSeek’s claim is that V4.1 Flash beats V4 Pro on “performance, cost, speed, and task completion time.” The published benchmarks mostly back that up, the weights are on Hugging Face under MIT, and the price is lower than V4 Flash was. The catch is the checkpoint: 552B parameters and about 510 GB on disk, three times V4 Flash, which changes the self-hosting math completely.

TL;DR

• Live since September 10, 2026 on the DeepSeek API as deepseek-flash; V4 Flash and the Flash vision preview are retired
• $0.15 in / $0.60 out per million tokens off-peak, $0.30 / $1.20 at peak, cache hits $0.003 off-peak and $0.006 at peak. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday; everything else, including weekends, is off-peak
• From September 14 at 12:00 Beijing time, deepseek-v4-pro requests route to V4.1 Flash at Flash prices. No V4.1 Pro date
• 552B total parameters, 8B active on prefill and 16B on decode, 1M context, 384K max output, image input built in, MIT license, ~510 GB FP8 checkpoint on Hugging Face
• On DeepSeek’s own table, V4.1 Flash beats V4 Pro on most agentic and coding benchmarks (DeepSWE v1.1 74.2 vs 62.7, Terminal-Bench 2.1 90.6 vs 87.9) and loses on a few knowledge ones (HLE 36.8 vs 42.7)
• Self-hosting floor is an 8-GPU node in practice, versus 2x H200 for V4 Flash. The KV cache is a quarter of V4 Flash’s per token, so the weights are the whole problem
• DeepSeek V4 Flash and V4 Pro are live on Yotta AI Gateway at flat rates with no peak windows; this post will note it when V4.1 Flash joins the catalog

Pricing

Three things worth noticing. V4.1 Flash is cheaper than V4 Flash was on every line, so the upgrade comes with a price cut, which DeepSeek has done before but rarely on the flagship-replacing release. The cache-hit price is 50x below the cache-miss price, so any workload with a stable system prompt or repeated document context should be built around caching first. And peak hours are Beijing business hours, which for a US team means roughly 9 pm to midnight and 2 am to 6 am Eastern on weekdays, so most US daytime traffic and all weekend traffic bills at the off-peak rate.

For V4 Pro users the change is larger. From September 14 the same request that cost $1.32 in and $3.96 out at peak costs $0.30 and $1.20, a 4.4x cut on input and 3.3x on output, because it’s being served by a different model. Whether that’s a bargain or a regression depends on your workload, which is the routing question below.

What happens to V4 Pro on September 14

DeepSeek’s notice is explicit: after 12:00 Beijing time on September 14, 2026, all requests to deepseek-v4-pro are routed to V4.1 Flash and billed at the V4.1 Flash price. Your code keeps working, the model name keeps resolving, and the model behind it changes.

If you run V4 Pro in production, do two things before Sunday. First, run your eval set against deepseek-flash today and compare it to your current V4 Pro outputs, because this is a different model with a different architecture, not a fine-tune, and system-prompt adherence tends to shift across releases like this. Second, decide whether to switch the model name yourself now rather than letting the router do it, so that the change lands on your schedule and shows up in your logs as a deliberate migration.

If the eval comes back worse on the tasks you care about, the fallback is self-hosting V4 Pro from the open weights, since the API will no longer serve it. That’s a 1.6T-parameter model, so it’s a real commitment. For most teams the practical answer will be to take the Flash routing, keep the eval running, and revisit when V4.1 Pro ships.

Specs

The architecture is new. V4.1 Flash is a causal encoder-decoder: a 20-layer causal encoder feeding a 20-layer decoder, which is why the active parameter count is different for prefill (8B) and decode (16B). DeepSeek pairs that with a second-generation compressed sparse attention that brings the persistent KV cache down to 890 bytes per token, roughly a quarter of V4 Flash and, with a sliding-window replay trick, about an eighth. A full 1M-token context costs well under a gigabyte of cache. That’s the memory win. The memory cost is 552B parameters that all have to live in VRAM, whatever the active count says.

Image input is native this time. V4 Flash was text-only with a separate vision preview; V4.1 Flash has a from-scratch vision encoder in the main model, and the retired deepseek-v4-flash-vision-exp name now points at it.

Benchmarks, with the usual caveat

Every number here is DeepSeek’s, from the model card, at maximum reasoning effort. There’s no independent replication yet; Artificial Analysis and the others will take a few days. Treat it as the vendor’s best case.

The pattern is clear and it matches the positioning. On agentic coding and terminal work, V4.1 Flash is a big step over V4 Flash and ahead of V4 Pro, and it lands next to GPT-5.6 Sol on DeepSWE and Terminal-Bench 2.1. On raw knowledge and reasoning without tools, GPQA and HLE, it’s a Flash-class model: slightly under V4 Flash on HLE, well under V4 Pro. Terminal-Bench 4.0 is the honest distance to the closed frontier, 31.2 against Sol’s 39.9 and GPT-6 Astra’s 57.9.

So “surpasses V4 Pro across all key metrics” is true for the workloads DeepSeek is optimizing for and not true for everything. If your V4 Pro usage is agent loops, code, and tool use, the routing on September 14 is probably an upgrade at a quarter of the price. If it’s long-form knowledge work with no tools, run the eval before you trust it.

Self-hosting: the floor moved up

This is the part that changes the most. DeepSeek V4 Flash fit on 2x H200: a 166.9 GB pre-compressed checkpoint with room for cache. V4.1 Flash ships as roughly 510 GB of FP8 safetensors across 48 shards. That is more than four H200s hold before overhead, and the KV cache savings don’t help, because at 890 bytes per token the cache was never going to be the constraint.

Working from the checkpoint size, the practical floor is an 8-GPU node. 8x H200 (1,128 GB) runs it with wide margin; 8x H100 80 GB (640 GB) fits the weights with about 130 GB left for activations and cache, which should work but is tighter than most teams want in production. 8x B200 or B300 is the comfortable choice. A 4x H200 pod (564 GB) is on paper enough for the weights alone and not something to plan around until someone publishes a working config. Those are estimates from the file sizes, not tested deployments; a full hardware breakdown will follow once the serving stacks settle.

The engines are ready: the model card lists vLLM, SGLang, Transformers, and Docker Model Runner, with recommended sampling of temperature 1.0, top_p 0.95 or 1.0, and max_tokens of 256K or more. DSpark speculative decoding is part of the model this time rather than a separate download.

The self-hosting calculus that made V4 Flash the easy choice, two GPUs for a frontier-adjacent model, is gone. V4.1 Flash is a node-scale model at a Flash-scale API price, which pushes most teams toward the API, and toward a gateway with flat pricing if the peak windows are a problem.

How to access it

On DeepSeek’s API, set the model name to deepseek-flash. The endpoint is OpenAI-compatible, so for most codebases that’s a one-line change; if you were calling deepseek-v4-flash you don’t have to change anything, though you should, so the name in your logs matches the model.

On Yotta AI Gateway, DeepSeek V4 Flash ($0.33 in / $0.99 out) and V4 Pro ($0.99 / $2.97) are live at flat per-token rates with no peak-hour windows and no cache tiers to manage, behind the same key as Qwen 3.8-Max, Qwen3.8-27B, GLM 5.3, and Kimi K3. V4.1 Flash is not in the catalog as of this writing; this post will note it when it is. Since the Gateway is OpenAI-compatible, an A/B between DeepSeek’s deepseek-flash and the Gateway’s V4 Flash on your own traffic is a base URL and a model string, which how to switch models without changing your code walks through.

Self-hosting: the weights are at deepseek-ai/DeepSeek-V4.1-Flash on Hugging Face under MIT. 8-GPU H200 and B300 nodes by the hour cover the floor described above, and the V4 Flash and Pro deployment tutorial is the closest starting point until a V4.1-specific one is published.

Frequently asked questions

What is DeepSeek V4.1 Flash? DeepSeek’s new default API model, released September 10, 2026. It’s a 552B-parameter mixture-of-experts model with a new encoder-decoder architecture, 1M context, native image input, and MIT-licensed open weights. It replaces V4 Flash and, from September 14, serves all V4 Pro requests.

How much does DeepSeek V4.1 Flash cost? $0.15 per million input tokens and $0.60 per million output off-peak; $0.30 and $1.20 at peak (01:00 to 04:00 and 06:00 to 10:00 UTC, weekdays). Cache hits are $0.003 off-peak and $0.006 at peak.

What is the model name for DeepSeek V4.1 Flash?deepseek-flash. The old deepseek-v4-flash and deepseek-v4-flash-vision-exp names still work but now point at V4.1 Flash.

What happens to DeepSeek V4 Pro? From 12:00 Beijing time on September 14, 2026, all deepseek-v4-pro requests are routed to V4.1 Flash and billed at V4.1 Flash rates. DeepSeek has not announced a V4.1 Pro date. The V4 Pro weights remain available for self-hosting.

Is DeepSeek V4.1 Flash better than V4 Pro? On DeepSeek’s published benchmarks, yes for agentic coding, terminal, and tool-use tasks (DeepSWE v1.1 74.2 vs 62.7, Terminal-Bench 2.1 90.6 vs 87.9), and no for knowledge-heavy reasoning without tools (HLE 36.8 vs 42.7, GPQA Diamond 90.9 vs 92.4). No independent numbers yet.

Is DeepSeek V4.1 Flash open source? Yes. Full weights are on Hugging Face under the MIT license, about 510 GB in FP8, with vLLM and SGLang support listed on the model card.

What hardware does DeepSeek V4.1 Flash need? Estimated from the checkpoint: an 8-GPU node. 8x H200 or 8x B200/B300 comfortably, 8x H100 80 GB tightly. That’s a big step up from V4 Flash’s 2x H200, because the model is three times larger on disk even though its KV cache is a quarter the size.

Is DeepSeek V4.1 Flash on Yotta AI Gateway? Not yet. DeepSeek V4 Flash and V4 Pro are live on Yotta AI Gateway at flat rates. This post will be updated when V4.1 Flash is added.

Bottom line

DeepSeek V4.1 Flash is a bigger model at a lower price that quietly retires both of its predecessors. For API users it’s a straight upgrade on agentic and coding work at a quarter of V4 Pro’s price, with a real regression risk on knowledge-heavy tasks that you should measure before September 14 rather than after. For self-hosters it ends the two-GPU era; this one needs a node.

The move that survives either outcome is the same one every launch this year has pointed at: keep the endpoint OpenAI-compatible, keep an eval set that runs in minutes, and treat the model name as configuration. Yotta AI Gateway has the current DeepSeek lineup and its open-weight rivals behind one key for exactly that reason, and 8-GPU nodes by the hour cover the self-hosted side when V4.1 Flash is worth owning.