AI ComparisonsOct 18, 20268 min read

DeepSeek V4.1 Flash vs V4 Pro: The Reroute That Never Came

DeepSeek V4.1 Flash vs V4 Pro: DeepSeek kept V4 Pro running after saying it would reroute it, yet Flash costs a quarter to a third as much and reads images.

On this page
  1. At a glance
  2. Is my deepseek-v4-pro call actually V4.1 Flash?
  3. What changed between V4 Pro and V4.1 Flash?
  4. How much cheaper is V4.1 Flash than V4 Pro?
  5. How did V4.1 Flash do on our decision tests?
  6. Which one should you self-host?
  7. Can you run DeepSeek V4.1 Flash or V4 Pro in Canada?
  8. Where V4.1 Flash falls short
  9. FAQ
  10. Sources

DeepSeek V4 Pro still runs. DeepSeek said in September it would send V4 Pro requests to the cheaper V4.1 Flash, then reversed the plan. As of 18 October 2026, its pricing page still serves deepseek-v4-pro as V4 Pro. Flash, updated 17 October, costs a quarter to a third as much and wins most of DeepSeek's tests.

If you call DeepSeek's API for agent or coding work, switch to `deepseek-flash`.

If your prompts are tuned for Pro and they work, keep Pro while you test Flash.

If you self-host, start with Flash. It has about a third of Pro's parameters, and only Flash reads images.

At a glance

DeepSeek V4.1 FlashDeepSeek V4 Pro
Released10 Sep 2026; updated 17 Oct 2026GA 13 Aug 2026
API model namedeepseek-flashdeepseek-v4-pro
Version served, 18 Oct 2026DeepSeek-V4.1-Flash-1017DeepSeek-V4-Pro-0813
Size552B parameters, 8B active reading input, 16B writing output1.6T parameters, 49B active
How much it can read at once1M tokens1M tokens
Image inputYesNo
LicenceMITMIT
Price per 1M tokens, peak (input / output)$0.30 / $1.20$1.32 / $3.96
Concurrent requests allowed2,500500

Sizes come from the V4.1 Flash and V4 Pro model cards. Prices and limits come from DeepSeek's pricing page as of 18 October 2026, in the Wayback Machine copy.

Is my deepseek-v4-pro call actually V4.1 Flash?

No. As of 18 October 2026, deepseek-v4-pro still runs V4 Pro, billed at Pro rates. DeepSeek withdrew the reroute before it took effect.

The sequence, from DeepSeek's own pages:

DateWhat DeepSeek published
10 Sep 2026The V4.1 Flash launch note says all deepseek-v4-pro requests "will route to V4.1-Flash at V4.1-Flash rates" from 04:00 UTC on 14 September, until V4.1 Pro launches.
12 Sep 2026PricingSaaS records the change on DeepSeek's pricing page: V4 Pro continues past 14 September.
Change log, 10 Sep entryDeepSeek's change log now reads: "In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged."
17 Oct 2026V4.1-Flash-1017 replaces the September Flash. Pricing and the model name stay the same.
18 Oct 2026The pricing page lists deepseek-v4-pro as DeepSeek-V4-Pro-0813 at its own prices.

The 10 September launch note still carries the original routing sentence, which is why many posts repeat it, including TechNode's launch report. The change log and pricing page are the later word.

Two names do route to Flash. The retired deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, and the pricing page says DeepSeek-V4.1-Flash-1017 answers them at the Flash price.

V4.1 Pro has not launched. DeepSeek promises notice before Pro changes again. Read the pricing page before you budget a quarter ahead.

What changed between V4 Pro and V4.1 Flash?

V4.1 Flash is a smaller model on a new design, and DeepSeek says it beats V4 Pro. It also reads images.

Three changes matter in practice:

  • Fewer parameters at work. Pro uses 49B of its 1.6T parameters on each token. Flash has 552B in total. It uses 8B to read your input and 16B to write. Less work per token makes it cheaper to serve.
  • Less memory for long inputs. A model keeps working memory for the text it has read, called the KV cache. DeepSeek says Flash's cache needs a quarter of the GPU memory of the previous generation. Cache hits are often most of an agent's bill.
  • It reads images. Flash takes screenshots and scanned pages. The pricing page lists vision as "Not supported" for Pro.

Here is how DeepSeek scores the two on the tests both announcements name. These are DeepSeek's own numbers, from the 13 August V4 Pro entry and the 10 September Flash entry in its change log, not independent results.

Task (DeepSeek-reported)V4.1 Flash, SepV4 Pro, Aug GA
Finishing multi-step jobs in a command-line terminal90.687.9
Very hard expert questions, no tools36.842.7
Very hard expert questions, allowed to search and run code63.960.0
Reproducing real software security flaws88.183.3
Building a whole code repository from a written spec65.461.5

Flash leads on four of five, by three to six points. Pro kept a six-point lead on hard questions answered without tools. The October update closes that gap: DeepSeek reports 42.6 for the new Flash against Pro's 42.7. A few points on a vendor's own run is close to the noise. Test both on your own work before you move a production job.

How much cheaper is V4.1 Flash than V4 Pro?

At list price, Flash costs 23% of Pro's rate for new input and 30% for output. Both halve their prices off-peak.

Per 1M tokensFlash, peakFlash, off-peakPro, peakPro, off-peak
Input, already cached$0.006$0.003$0.044$0.022
Input, new$0.30$0.15$1.32$0.66
Output$1.20$0.60$3.96$1.98

A token is about three quarters of a word. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else is off-peak, which covers the whole working day in Toronto.

A worked example: a month of agent work that sends 100M new input tokens and gets 20M output tokens back.

Same month of workV4.1 FlashV4 Pro
All at peak rates$54.00$211.20
All off-peak$27.00$105.60

The example ignores cache hits, which cut both bills further. Flash also allows 2,500 concurrent requests against V4 Pro's 500, which matters more than price once several agents run at once.

Our earlier V4 Pro comparison quoted lower V4 Pro prices from a third-party router in August. The table above is DeepSeek's own list price as of 18 October 2026.

How did V4.1 Flash do on our decision tests?

On DecideBench, our public test of 400 decisions such as whether a refund falls inside a return window, V4.1 Flash got 99.2% right. It cost $455.09 per million decisions on Together AI, with a median wait of 356 ms. The run covered 28 September to 9 October 2026, before the 17 October update.

The scope is narrow. DecideBench measures picking one answer from a fixed list. It says nothing about coding or writing, and we did not run V4 Pro. In the same run, the older DeepSeek V4 Flash scored 99.8% at $197.28, so on this task the new model was faster but not more accurate. The full table is in decision model vs LLM.

Which one should you self-host?

Self-host Flash unless your own tests show Pro doing something Flash cannot. Both use the MIT licence, which allows commercial use and changes, so size and features decide it.

Self-hostingV4.1 FlashV4 Pro
Total parameters552B1.6T
Weights on Hugging FaceDeepSeek-V4.1-Flash-1017DeepSeek-V4-Pro
Engines the card namesvLLM, SGLang, Docker Model RunnerInference code in the repo
Image inputYesNo

Neither card gives a GPU count or memory requirement. Your workload sets the hardware: how long inputs run and how many requests arrive at once. Flash's smaller cache helps most on long inputs. Pro's card asks for at least 384K tokens of context in its deepest reasoning mode.

The October Flash has its own repository. The September weights stay up under the original V4.1 Flash name.

Can you run DeepSeek V4.1 Flash or V4 Pro in Canada?

Yes, both. Anyone can download the weights, so either model can run on dedicated GPUs in a Canadian cloud region or on your own hardware. Prompts and outputs then stay in the country.

DeepSeek's own API is a different matter. Its privacy policy says it collects, processes and stores personal data in the People's Republic of China. Our post on sending code to overseas LLM APIs covers what that means for proprietary code.

Running it yourself in Canada means:

  • Download. Both sets of weights are on Hugging Face under MIT. Flash is the smaller of the two.
  • Where. AWS ca-central-1 in Montreal or ca-west-1 in Calgary, or your own racks. The Canadian residency guide maps the other routes.
  • Trade-offs. Large GPU instances are scarcer in Canadian regions than in US ones. You pay for GPUs while they sit idle. You own the serving, patching and safety filtering.

XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account, or on-premises, so prompts and outputs stay in the country, then builds the workflow on top. The team runs the work and the model takes part. How we deploy covers the build.

If you need this model's capability with inference kept in Canada, bring one job to a free 30-minute call.

Where V4.1 Flash falls short

  • Hard questions without tools. The September Flash trailed Pro here. The October update closes the gap on paper only.
  • Vendor numbers only. Every coding score here is DeepSeek's. We have not tested either model on coding.
  • Moving targets. DeepSeek has changed Pro's future once already. V4.1 Pro has no date.
  • What we did not test. Our only number is the decision test above, on the September Flash.

For the open models V4 Pro competed with in August, see the best open-source coding models of 2026.

FAQ

Does deepseek-v4-pro route to V4.1 Flash?

No. DeepSeek announced on 10 September 2026 that deepseek-v4-pro calls would go to V4.1 Flash from 14 September, then reversed the plan. Its change log says V4 Pro continues "with the billing method remaining unchanged". As of 18 October 2026 the pricing page lists deepseek-v4-pro as DeepSeek-V4-Pro-0813. Only the retired deepseek-v4-flash names route to Flash.

Is DeepSeek V4.1 Flash better than V4 Pro for coding?

On DeepSeek's own tests, mostly yes. The September Flash beat V4 Pro on terminal tasks, building repositories from a spec and reproducing security flaws, by three to six points. Those are vendor-reported scores, and we have not run either model on coding. Try both on a sample of your own tickets before you switch a production agent.

How much does DeepSeek V4.1 Flash cost?

At peak, $0.30 per million new input tokens, $0.006 per million cached input tokens and $1.20 per million output tokens. Off-peak rates are half. That is about a quarter of V4 Pro's input price and under a third of its output price, per DeepSeek's pricing page as of 18 October 2026.

Can I run DeepSeek V4.1 Flash locally?

Yes. The October weights are on Hugging Face as DeepSeek-V4.1-Flash-1017, under the MIT licence. The model card names vLLM, SGLang and Docker Model Runner, and says quantized versions work in llama.cpp, Ollama and LM Studio. It gives no GPU count, so size the hardware from your input lengths and traffic.

When is DeepSeek V4.1 Pro coming out?

DeepSeek has not given a date. Its 10 September 2026 launch note names V4.1 Pro as the successor to V4 Pro, and nothing on its news or change log pages had announced it as of 18 October 2026.

Sources

Written by

Cho Yin Yong

Principal AI Solutions Engineer, XY Space

Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.

More from Cho Yin Yong

Share this article

Work with us

We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.

Start a project
Work with us

Book a call.We'll come back with specifics.

Start with the map of your organization, or with the one job that hurts. Measured in hours and money, and everything we build stays yours.

Loading form…