DeepSeek V4.1 Flash Launches — and V4 Pro Retires on Sept 14

DeepSeek released V4.1 Flash on 10 September 2026, and the part that matters most to anyone running it in production is not the benchmark chart. It is a date: on 14 September at 12:00 Beijing time, requests to deepseek-v4-pro start being served by V4.1 Flash and billed at V4.1 Flash rates. If you pinned the Pro model name anywhere, you have four days.

This is a same-day report based on DeepSeek's announcement as carried by Chinese tech media. Some details — particularly international pricing — are still settling; see the note at the bottom before you act on the numbers.

1. What actually launched

V4.1 Flash is the smallest model in DeepSeek's new architecture family. The design goals stated for the family are a higher capability ceiling, faster inference, greater throughput, and the ability to scale to larger parameter counts later.

Two things distinguish it from a routine point release:

It is natively multimodal. Vision understanding is built in rather than bolted on through a separate experimental model. Previously, vision lived in a separate DeepSeek-V4-Flash-Vision-Exp model; that separation is now gone.

The consumer app modes merged. DeepSeek folded its separate Fast, Expert, and Image modes into one unified mode that detects query complexity, activates vision when it sees an image, and scales reasoning effort to the task. Users no longer switch manually.

2. The asymmetric architecture

This is the technically interesting part, and DeepSeek disclosed it for the first time with this release.

V4.1 Flash is a 552B-parameter mixture-of-experts model using a new Causal-Encoder-Decoder structure. The key point is that the two ends are not symmetric:

Stage	Active parameters
Input / encoding	8B
Output / decoding	16B

Not every parameter is touched on every inference pass. The idea is to spend different amounts of compute on different stages of the job rather than running the same intensity end to end — which is how a model this size ends up cheaper to serve than its parameter count suggests.

DeepSeek also reports a new pre-training approach and a larger-scale reinforcement learning post-training phase.

The KV Cache compression matters more than it sounds

Against the previous generation, V4.1 Flash needs roughly one quarter the HBM and one eighth the SSD for KV Cache. Against DeepSeek's first-generation models, the company puts the cumulative reduction at 437×.

Why this shows up on your bill: in agent workloads, cache hits are a large share of input tokens. Shrinking the cache is a direct cut to the cost of long, repetitive context — the exact shape agent loops have. This is paired with the cached-input price cut below, and the two reinforce each other.

3. Official specs and pricing

From DeepSeek's Models & Pricing page (USD, per 1M tokens):

	V4.1 Flash (deepseek-flash)	V4 Pro (deepseek-v4-pro)
Context length	1M	1M
Max output	384K	384K
Cached input (off-peak)	$0.003	$0.022
Cached input (peak)	$0.006	$0.044
Uncached input (off-peak)	$0.15	$0.66
Uncached input (peak)	$0.30	$1.32
Output (off-peak)	$0.60	$1.98
Output (peak)	$1.20	$3.96
Concurrency limit	2500	500
Vision	Yes	No

Off-peak rates are half of peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday — everything else, including all weekend, is off-peak. That off-peak window remains the single biggest lever most people never pull.

The headline change is cached input down 60% versus the previous pricing. Combined with the KV Cache work above, this is a meaningful cut for anyone running long-context or agentic workloads rather than one-shot prompts.

What the forced migration actually saves you

Because V4 Pro requests get billed at V4.1 Flash rates from 14 September, you can price the change before it happens. At off-peak, output drops from $1.98 to $0.60 per million tokens — roughly a 70% cut — and unfavourable as the trigger is, the concurrency limit also rises from 500 to 2500.

4. Migration: what to change, and by when

This is the section worth acting on.

Switch to the new model name. V4.1 Flash is live on the DeepSeek API under deepseek-flash. Change the model field and you are on it.

Old Flash names are retired but routed. deepseek-v4-flash and deepseek-v4-flash-vision-exp are taken down. For compatibility, requests using those names are temporarily routed to V4.1 Flash. Routing is a safety net, not a plan — the mapping can change, and you lose the ability to reason about which model actually answered.

V4 Pro has a hard date. From 12:00 Beijing time on 14 September 2026, until a V4.1 Pro is released, requests to deepseek-v4-pro are all served by V4.1 Flash and billed at V4.1 Flash rates. That is both cheaper and, per DeepSeek's own numbers, better — but it is a silent behaviour change if you have not read this.

If you pin model names in config, environment variables, or a database, grep for deepseek-v4-pro and deepseek-v4-flash this week.

5. Where it does not close the gap

The honest read, because "beats our own last flagship" is a narrower claim than it sounds.

DeepSeek says V4.1 Flash tops the chart on software engineering, security, and automation benchmarks — DeepSWE v1.1, CyberGym, and Automation-Bench — and exceeds V4 Pro overall on capability.

But on GPQA Diamond, the graduate-level cross-disciplinary science reasoning benchmark, it still trails leading closed models including GPT-5.6 Sol and Claude Opus 5.

So: very strong on agentic and coding work per unit cost, not yet the top model for hard scientific reasoning. If your workload is the latter, this release does not change your answer.

6. FAQ

Do I have to migrate off V4 Pro?

Not strictly — requests keep working. But from 14 September they are answered by V4.1 Flash and billed at Flash rates. If you were on V4 Pro deliberately, that decision stops being honoured on that date.

What is the new model name?

deepseek-flash. The old deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired, with temporary routing for compatibility.

Is V4.1 Flash multimodal?

Yes, natively. Vision understanding is built into the model; you no longer route image input to a separate vision model.

How much cheaper is it?

Cached input dropped 60%. Off-peak output is $0.60 per million tokens against $1.98 on V4 Pro, and the concurrency limit goes from 500 to 2500. Peak is exactly double off-peak on all three tiers.

Is a V4.1 Pro coming?

DeepSeek refers to a future V4.1 Pro, which is what the V4 Pro routing is described as a stopgap until. No date has been announced.

Does this affect off-peak pricing strategy?

It strengthens it. Peak is 2× off-peak on all three tiers, and cached input is now much cheaper — so shifting batch work to off-peak compounds with the cache discount. See off-peak pricing for how the windows work.

Sources

Reported 10 Sep 2026.

Primary source (used for all specs and pricing in section 3):

• DeepSeek API Docs — Models & Pricing — official model names, versions, 1M context / 384K max output, USD prices for every tier, concurrency limits, feature support, the definition of peak hours, and the V4 Pro retirement and routing rule.

Announcement details as carried by Chinese tech media (architecture, benchmarks, timing):

• 第一财经 / Yicai via Sina Finance — 「DeepSeek V4.1 Flash 转正，模型价格战继续」 — 552B 参数、Causal-Encoder-Decoder 非对称结构（输入 8B / 输出 16B）、KV Cache 压缩（HBM 1/4、SSD 1/8、累计 437×）、基准测试（DeepSWE v1.1 / CyberGym / Automation-Bench 领先，GPQA Diamond 仍有差距）、9 月 10 日 12:00 生效定价（闲时缓存命中 0.02 元 / 未命中 1 元 / 输出 4 元，高峰 2 倍，缓存命中降价 60%）、V4 Pro 9 月 14 日 12:00 起路由。
• TechWeb via Sina Tech — 「DeepSeek V4.1 Flash 正式发布：性能超 V4 Pro，API 定价同步下调」 — 模型名 deepseek-flash、旧模型下线但临时路由、deepseek-v4-pro 路由与计费规则、峰谷定价机制。
• IT之家 — 快速/专家/识图模式合并为统一智能模式 — V4.1 Flash 发布时间、V4 Pro 服务 9 月 14 日下线。

Caveats, stated plainly. The USD figures in section 3 are from DeepSeek's own pricing page and are the authoritative ones. Benchmark results, including the KV Cache reduction figures, are DeepSeek's own and were not independently reproduced. Prices can change — DeepSeek reserves the right to adjust them, and we recommend checking the official page before recalculating a budget. Check the verified date at the top of this page.

We have no affiliation with DeepSeek and carry no affiliate links on this page.