DeepSeek V4.1 Flash: second Flash in a row — 400 tok/s, native multimodality, and the Sept 10 price cut as a cost artifact

DeepSeek V4.1 Flash: second Flash in a row — 400 tok/s, native multimodality, and the Sept 10 price cut as a cost artifact

TL;DR: On September 10, 2026, DeepSeek shipped V4.1 Flash — the second Flash model in one week, again a temporary build with a stable successor. For builders this is not a ranking story but arithmetic: prices per million tokens dropped clearly, the old Pro plan redirects to the Flash tier starting September 14, and every cached workload benefits. If you pin fixed model IDs, start tracking the date of the next stable identifier.

The essentials at a glance

DeepSeek-V4.1-Flash is the successor in the Flash stack: same model, more tokens per dollar, and the Pro plan gets the same price via the redirect. Anyone pinning deepseek-flash captures the savings automatically.

The two price tiers in detail

The off-peak rate is exactly half the peak rate. DeepSeek defines peak as 01:00–04:00 UTC and 06:00–10:00 UTC, Monday to Friday. In numbers:

For a simple 10K-token cache miss plus 2K output: about $0.03 at peak, $0.015 off-peak. The old V4 table (August 2026) was still at $0.22/$0.44 for cache misses and $0.66/$1.32 for output. So the savings apply not only to new but also to existing cached workloads.

The redirect mechanic from September 14

The decisive lever is not just the new price list but what it does to the Pro plan. From September 14, 2026, 04:00 UTC (12:00 Beijing), DeepSeek routes every request to deepseek-v4-pro automatically to V4.1-Flash. The Pro plan effectively disappears behind the Flash price:

• Cache-miss input: from $0.66 to $0.15 — roughly 77% cheaper.
• Output: from $1.98 to $0.60 — roughly 70% cheaper.

The models deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired and redirect to deepseek-flash. A builder with pinned IDs thus sees four model names changing behind them without touching any code.

Cost per task: the numeric artifact

The easiest way to see the effect is a small calculation. Example: a support task with 10,000 tokens of cache-miss input and 2,000 tokens of output.

With two 1-million-token batches — one at peak, one off-peak — the savings add up to about $17 per million. That is the number that belongs in the decision table.

What it means for your stack

The release notes describe V4.1-Flash as a small model of the new architecture family with a native visual understanding unit; official DeepSeek tables list 221–235 tok/s, independent runs such as WorldofAI report up to 400 tok/s depending on hardware. The 1M-token context is retained, max output is 384K.

The decision rule for this week:

• Pinning deepseek-v4-pro? From September 14 you automatically get the Flash price. Check: are old Pro cost assumptions still in your code?
• Pinning deepseek-flash? You save immediately. Check: is the cache-hit ratio high enough for the cache-hit rate to apply?
• Comparing both models: compute the tiers separately — 01:00–04:00 and 06:00–10:00 UTC are expensive, everything else is half price.

Multi-model setups: deepseek-v4-flash and deepseek-v4-flash-vision-exp no longer exist as separate IDs — both names redirect to deepseek-flash. Hardcoded model lists will show one line less than expected.

Local models: the V4.1 family is a 552B MoE with 8B active parameters on input and 16B on output, MXFP4, roughly 510 GB disk footprint for the full stack, licensed under MIT. The local inference option remains; the question is your GPU slot.

FAQ

When exactly did the new Flash price take effect? The prices became active with the V4.1-Flash release on September 10, 2026, 04:00 UTC. From that point, deepseek-flash is the model name in the API. The old names have since run as redirects to the new build.

What happens to the Pro plan after September 14? The Pro plan is automatically redirected to V4.1-Flash and billed at Flash prices. Concretely: 77% lower cost for cache-miss input and roughly 70% for output. A new Pro generation arrives only when V4.1 Pro is released.

How are peak and off-peak defined? Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays. All other hours are off-peak, priced at exactly half. A simple lever: move batch processing into off-peak windows and run only latency-critical requests at peak.

Why are there two different throughput figures? Official DeepSeek tables list 221–235 tok/s for V4.1-Flash. Third-party runs such as the WorldofAI test report up to 400 tok/s depending on hardware and prompt size. For cost calculations use the official range; the 400 figure is the best case of a single test run.

How does the model relate to local GGUF setups? The GGUF variant of Qwen3.8-27B at 11.8 GB remains a separate point in the same price landscape: local stacks are capped to the 16 GB slot, and API prices are now the anchor. If you combine both, write the off-peak rule directly into the scheduling table for API runs.

Sources

• DeepSeek docs and changelog: https://api.deepseek.com/docs
• Apidog price overview (peak/off-peak tables): https://apidog.com/blog/deepseek-v4-1-flash-pricing
• Coursiv — V4.1 Flash metrics and redirect: https://coursiv.io/blog/deepseek-v4-1-flash
• CodingFleet — architecture, context, max output: https://codingfleet.com/blog/deepseek-v4-1-flash-vs-claude-opus-5
• Eesel — release summary: https://www.eesel.ai/blog/deepseek-v4-1-flash
• WorldofAI test log: https://www.youtube.com/watch?v=T2dnchLabZQ