So the date that matters for your plan is 5 November, not 19 November. Test and switch before capacity starts moving. Once you have switched, the retirement on 19 November changes nothing for you.
DeepSeek V4 Pro is being retired: what do I need to change, by when, and will my workload still work on V4.1 Flash? You probably got our email of 29 October 2026 and have a product or agent that calls V4 Pro today. This guide gives the dates, the reason we recommend V4.1 Flash, the code change and the checks to run first.
What changes on 5 and 19 November
DeepSeek V4 Pro changes in two steps on our serverless inference, the API where you call open models and pay per token. Our email of 29 October 2026 sets out the dates.
| Date | What happens to DeepSeek V4 Pro | What you should do |
|---|---|---|
| Until 5 November 2026 | Nothing changes. | Test V4.1 Flash on your own prompts. |
| From 5 November 2026 | We move capacity to V4.1 Flash, so V4 Pro may be slower. | Switch your production traffic. |
| From 19 November 2026 | V4 Pro is retired. We will not switch your model automatically. | Nothing should still call V4 Pro. |
"Retired" means we stop serving the model. A request that still names it will not be answered by V4 Pro, and we do not forward it to another model. That is deliberate. A silent switch would change your outputs and your bill without your knowledge.
The two-week gap is your test window. A workload that runs late, or one that waits for long answers, will feel reduced capacity first. Switching before 5 November keeps your speed predictable.
Why we recommend DeepSeek V4.1 Flash
DeepSeek V4.1 Flash is the replacement we recommend, and our email calls it "great for agentic coding". Agentic coding means a model that edits code over many steps and calls tools, such as a test runner, along the way.
DeepSeek itself points the same way. Its release note says tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime. DeepSeek is phasing out V4-Pro too. On DeepSeek's own API, requests for deepseek-v4-pro have gone to V4.1-Flash since 14 September 2026. That rerouting happens on DeepSeek's API only; on ours, nothing is rerouted. Those are DeepSeek's results, not ours, so test your own prompts.
The two models are built differently. Both are mixture-of-experts models: only part of the network works on each token, which keeps each token cheaper to compute. DeepSeek V4.1 Flash has 552B parameters, and at most 16B of them work on any one token. DeepSeek V4 Pro 0813, the August 2026 release of the V4 Pro we are retiring, has 1.6T parameters with 49B active. DeepSeek publishes both under the MIT licence.
Both models run on our EU-hosted infrastructure, so switching does not change where your requests are processed. We do not keep your prompts or outputs, and we never use them for training.
We also measure speed ourselves. Time to first token is the wait before the answer starts. In our status checks, V4.1 Flash had a median of 256 ms. That is the median over the seven days to 30 October 2026. Its median output speed was 87 tokens per second. Both figures come from short requests and change from week to week; the status page shows availability live.
What the switch does to your bill: a worked example
The switch lowers every token price on our model pages. Prices are in US dollars per 1M tokens, read on 30 October 2026.
| Token type | DeepSeek V4 Pro 0813 | DeepSeek V4.1 Flash | Change |
|---|---|---|---|
| Input | $2.00 | $0.50 | 75% less |
| Cached input | $0.50 | $0.13 | 74% less |
| Output | $4.00 | $1.50 | 62.5% less |
Cached input is the part of a prompt that repeats the start of an earlier request. We bill it at the lower rate. Our email quotes the 75% and 62.5% figures, and the table shows where they come from.
Take a hypothetical coding agent. In one month it sends 100M new input tokens and 200M cached input tokens, and it writes 15M output tokens. Agents resend their history on every step, so cached input is often their biggest line.
Hypothetical month: 100M input, 200M cached input, 15M output tokens
DeepSeek V4 Pro 0813: 100 x $2.00 = $200.00
200 x $0.50 = $100.00
15 x $4.00 = $60.00
total $360.00
DeepSeek V4.1 Flash: 100 x $0.50 = $50.00
200 x $0.13 = $26.00
15 x $1.50 = $22.50
total $98.50
Difference: $360.00 - $98.50 = $261.50 per month (72.6% less)The example uses the same token counts for both models, and that favours V4.1 Flash. It may write more tokens for the same task, which would shrink the saving. Reasoning tokens, the model's working before its answer, are billed as output. The cached rate is not guaranteed either: our page calls caching best effort. Read your real token counts for a week after the switch.
Change the model ID
The change is one value: the model ID, the name you put in the model field of each request. Our API stays the same OpenAI-compatible endpoint, so your SDK, base URL and API key do not change. Lyceum is a sovereign European AI inference provider for open-source models, and this endpoint runs on GPUs in European data centres.
Find every place that names the old model first. Config files and environment variables are easy to miss. A search like this one, run from your repository root, is a starting point.
grep -rn "deepseek-v4-pro" .Then point the OpenAI SDK at our base URL with your Lyceum API key. Set the new model ID.
from openai import OpenAI
client = OpenAI(
base_url="https://api.lyceum.technology/openai/v1",
api_key="YOUR_LYCEUM_API_KEY",
)
response = client.chat.completions.create(
model="deepseek/deepseek-v4.1-flash",
messages=[{"role": "user", "content": "Summarise this diff in three bullet points."}],
)
print(response.choices[0].message.content)Keep the model ID in one config value, not in every call. The next switch is then a one-line change.
Five behaviour checks before 5 November
Five behaviours of V4.1 Flash decide whether your code keeps working: reasoning, output limits, tool calling, caching and images. Each comes from our DeepSeek V4.1 Flash page, so test it with your own prompts before you move production traffic.
| Check | Why it matters | What to do |
|---|---|---|
| Reasoning | Reasoning is on by default, and the reasoning trace (the model's working) counts as output tokens. | Turn it off with reasoning_effort: "none" where you do not need it, and set max_tokens high enough for working plus answer. |
| Output limits | One request can write up to 131,072 tokens and run up to 300 seconds. The context window, the most text one request can hold, is 1M tokens. | Set stream: true for long answers. |
| Tool calling | Tools work in the OpenAI format, but tool_choice: "required" or a named function is not enforced. | Check message.tool_calls and fall back when it is empty. |
| Prompt caching | Prompt caching bills a repeated prompt start at the cached rate. It is automatic and best effort. | Keep the stable part of your prompt first, and read usage.prompt_tokens_details.cached_tokens. |
| Images | V4.1 Flash reads images sent as image_url content parts. | Test your image inputs if your workload sends any. |
Tool calling deserves the most care in an agent. If your code assumes a tool call always comes back, a plain-text answer can stall the loop. Our guide on what breaks when switching models covers this in depth.
How we retire models from now on
We retire models in the same three steps whenever we can, as our email of 29 October sets out. You will see this order again.
- We email you with the model we recommend. Nothing changes yet.
- We reduce capacity, so the model may be slower.
- The model is retired.
Step 2 is your real deadline each time. Treat the first email as the start of your test window, not as a notice to file. Our article on model deprecation risk shows how to plan for this across providers.
Need help with this switch? Reply to our email with a short description of your workload, and we will help you move it. If your tests pass, open the DeepSeek V4.1 Flash model page, copy the model ID and run your test prompts against it before 5 November.
Frequently asked questions
Will my requests to DeepSeek V4 Pro be moved to V4.1 Flash automatically?
No. From 19 November 2026 V4 Pro is retired on our API, and we do not switch your model for you. You change the model name yourself.
Does switching change where my data is processed?
No. V4 Pro 0813 and V4.1 Flash both run on our EU-hosted infrastructure, so the place of processing stays the same.
What if V4.1 Flash does not work for my workload?
Reply to our email with your workload before 5 November. We will look at it with you and help you switch.