DeepSeek V4.1 Flash Guide

Official pricing (re-checked 2026-10-05 — unchanged)

USD per 1M tokens. Off-peak rates are half the peak rates.

📘 cache hit vs cache miss: if the beginning of your prompt is still in DeepSeek's cache it counts as a cache hit (input $0.003 per 1M tokens — 1/50 of cache miss), otherwise it counts as a cache miss ($0.15). Sending the same long system prompt in the same order makes hits more likely.

Unit price vs V4 Pro (current rates)

⏰ Peak hours: 01:00-04:00 and 06:00-10:00 UTC, Monday-Friday only (excluding Chinese public holidays) (JST 10:00-13:00 / 15:00-19:00; Beijing 09:00-12:00 / 14:00-18:00). Weekends and Chinese public holidays are off-peak all day. The peak/off-peak scheme took effect 2026-08-16 16:00 UTC, and footnote (2) on the official pricing page reads "Monday through Friday (excluding Chinese holidays)". Source: api-docs.deepseek.com (first read 2026-09-11; footnote captured 2026-09-23, re-checked 2026-10-05).

Estimate your own cost with the token calculator, which lists both off-peak and peak rows for V4.1 Flash. For cross-provider comparison see pricing (DeepSeek).

V4 Pro keeps running (as of 2026-09-13)

Only what DeepSeek's official documentation states.

Model names at a glance (as of 2026-09-13)

Source: Models & Pricing notes (1) and (2) · Change Log (retrieved 2026-09-13). See the V4 Pro continuation page for the full picture.

Specifications (official feature table)

Source: api-docs.deepseek.com Models & Pricing (retrieved 2026-09-11). See DeepSeek's Rate Limit & Isolation page for concurrency details.

Official benchmarks (published 2026-09-10)

Numbers as published by DeepSeek in its change log. Not independently re-measured by this site.

Source: Change Log "DeepSeek-V4.1-Flash Release" (2026-09-10). These are vendor-reported figures.

Setup in VS Code GitHub Copilot Chat (3 steps)

The steps the operator (okamo channel) actually used with the V4 family. Extension names and model-picker labels can change, so check the official docs too.

💡 If you set the model name in code, use deepseek-flash. Code that still points at deepseek-v4-pro keeps working as-is (V4 Pro continues; see above).

Operator's setup record (2026-07-04, at the time of V4 Pro): okamo channel notes

Running it on Microsoft Foundry (Azure) — deploying DeepSeek-V4.1-Flash with the CLI (measured 2026-09-28)

DeepSeek's first-party API is not the only route: you can run DeepSeek models through Microsoft Foundry on Azure. In a measurement on 2026-09-28, we deployed the current model DeepSeek-V4.1-Flash (version 2026-09-10) with the az CLI and got a response back from the OpenAI-compatible endpoint. Below: ① availability ② deploy ③ choosing capacity ④ verifying the call ⑤ the pitfalls, using the exact commands we ran.

Prerequisites

• An Azure subscription where you can create resources (registering a resource provider requires Owner)
• The Azure CLI (az). The az cognitiveservices commands ship with core, so Azure CLI 2.60+ needs no extra extension
• You are signed in to the target subscription with az login

Placeholders used below

Step 0 — Prepare the Azure CLI and register the resource provider

Doing this first avoids the MissingSubscriptionRegistration error later.

# Check the Azure CLI version (2.60+ recommended)
az version

# Sign in if needed
az login

# Select the subscription to use
az account set --subscription "<SUBSCRIPTION_ID_OR_NAME>"

# Register the resource provider (avoids MissingSubscriptionRegistration)
az provider register --namespace Microsoft.CognitiveServices

# Check registration state (can take a few minutes to become Succeeded)
az provider show --namespace Microsoft.CognitiveServices --query registrationState -o tsv

Step 1 — Create a resource group

az group create \
  --name "<RESOURCE_GROUP>" \
  --location "<LOCATION>"

Step 2 — Create the Azure AI Services account

With --custom-domain, the endpoint becomes https://<ACCOUNT_NAME>.openai.azure.com/.

az cognitiveservices account create \
  --name "<ACCOUNT_NAME>" \
  --resource-group "<RESOURCE_GROUP>" \
  --kind AIServices \
  --sku S0 \
  --location "<LOCATION>" \
  --custom-domain "<ACCOUNT_NAME>" \
  --assign-identity \
  --allow-project-management true \
  --yes

Step 3 — Deploy DeepSeek-V4.1-Flash

List the models actually available to your account, then deploy with those values. Checking the real thing first is how you stay safe while the docs lag behind.

# List deployable DeepSeek models and their versions
az cognitiveservices account list-models \
  -n "<ACCOUNT_NAME>" -g "<RESOURCE_GROUP>" \
  --query "[?contains(name,'DeepSeek')].{name:name,format:format,version:version}" -o json

# Deploy (SKU is GlobalStandard; see the notes below for --sku-capacity)
az cognitiveservices account deployment create \
  --name "<ACCOUNT_NAME>" \
  --resource-group "<RESOURCE_GROUP>" \
  --deployment-name "<DEPLOYMENT_NAME>" \
  --model-name "DeepSeek-V4.1-Flash" \
  --model-version "<MODEL_VERSION>" \
  --model-format "DeepSeek" \
  --sku-name "GlobalStandard" \
  --sku-capacity 20

# Confirm the deployment (model name, SKU, capacity)
az cognitiveservices account list-deployments \
  -n "<ACCOUNT_NAME>" -g "<RESOURCE_GROUP>" \
  --query "[].{name:name,model:properties.model.name,version:properties.model.version,sku:sku}" -o json

💡 On 2026-09-28, list-models returned DeepSeek-V4.1-Flash (version 2026-09-10, format DeepSeek) and the deployment went through with those values. Version strings can differ by environment, so don't paste the value from this article — use what list-models returns.

# Check current quota usage and limits (LOCATION = your deployment region)
az cognitiveservices usage list --location "<LOCATION>" -o table

Step 4 — Export the connection settings

# Endpoint (OpenAI-compatible) and API key
export AZURE_AI_ENDPOINT="https://<ACCOUNT_NAME>.openai.azure.com/openai/v1/"
export AZURE_AI_API_KEY="$(az cognitiveservices account keys list \
  --name "<ACCOUNT_NAME>" \
  --resource-group "<RESOURCE_GROUP>" \
  --query key1 -o tsv)"

# The API's model field takes the deployment name
export DEPLOYMENT_NAME="<DEPLOYMENT_NAME>"

Step 5 — Verify with curl (and read the rate-limit headers)

curl -sS -D - "$AZURE_AI_ENDPOINT/chat/completions" \
  -H "Content-Type: application/json" \
  -H "api-key: $AZURE_AI_API_KEY" \
  -X POST \
  -d '{
    "model": "'"$DEPLOYMENT_NAME"'",
    "messages": [
      { "role": "user", "content": "Reply with just OK." }
    ]
  }'

✅ A JSON response means you are connected. Note that model takes the deployment name (passing the model name DeepSeek-V4.1-Flash can return a 404). -D - prints the rate-limit headers (x-ratelimit-limit-requests / x-ratelimit-limit-tokens), so you can confirm your capacity is in effect. If you prefer not to keep keys in environment variables, pass -H "api-key: <API_KEY>" directly. For Entra ID (formerly Azure AD) auth, the official tutorial's approach with DefaultAzureCredential and scope https://ai.azure.com/.default also works.

MissingSubscriptionRegistration
The subscription is not registered to use namespace 'Microsoft.CognitiveServices'.

First-party API vs the Azure route (as of 2026-10-05)

Same model, not the same service. Only the differences that matter in operations:

For the data-handling side in detail, see the zero data retention explainer and the China AI section of the China AI page.

Sources: Microsoft Foundry model catalog — DeepSeek-V4.1-Flash (version 2026-09-10, Preview, context window 1000k), Foundry blog — DeepSeek-V4.1-Flash is coming to Microsoft Foundry (updated 2026-09-25; two paths and the price table), az cognitiveservices account deployment (ms.date 2026-09-01), Foundry Models quotas and limits (TPM/RPM and 429 behaviour), Deployment types and Data, privacy, and security (data location and training). The command sequence and capacity figures were measured by the operator (okamoちゃんねる) on 2026-09-28 and re-checked against primary sources retrieved 2026-09-29.

FAQ

Q. How much does V4.1 Flash cost?

A. Per 1M tokens: input (cache miss) $0.15 off-peak / $0.30 peak, input (cache hit) $0.003 off-peak / $0.006 peak, output $0.60 off-peak / $1.20 peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays only (JST 10:00-13:00 / 15:00-19:00); weekends are off-peak all day. Off-peak is half the peak rate.

Q. What happens if I keep calling deepseek-v4-pro after September 14?

A. It keeps working. DeepSeek states that, in response to user demand, it will continue providing API services for DeepSeek V4 Pro after September 14, 2026, with billing unchanged. The earlier announcement that requests would route to V4.1 Flash from 9/14 has been withdrawn. No code or API-key change is required.

Q. When will DeepSeek V4.1 Pro be released?

A. No release date has been announced as of 2026-09-13. The earlier note that the routing would last "until the future release of V4.1 Pro" has itself been replaced by the V4 Pro continuation notice. This site will update as soon as an official announcement appears.

Q. Can it read images?

A. Yes. V4.1 Flash is built on the new architecture family with native multimodal visual understanding, and the official feature table marks Vision as supported (V4 Pro did not). For a cross-model view of image, audio, video and PDF support, see the multimodal LLM guide.

Q. Can I keep using deepseek-v4-flash?

A. Those names are still accepted for compatibility, but the underlying models are retired: requests are served by V4.1 Flash and billed at the V4.1 Flash price. New code should use deepseek-flash.

Q. Context length and max output?

A. 1M-token context and 384K-token maximum output. Thinking mode supports both non-thinking and thinking, with thinking as the default.

Q. Can I run DeepSeek V4.1 Flash on Azure (Microsoft Foundry)?

A. Yes. In a measurement on 2026-09-28 we deployed DeepSeek-V4.1-Flash (version 2026-09-10) with az cognitiveservices account deployment create and got responses from the OpenAI-compatible endpoint. It is Preview (public preview), and as of 2026-10-05 Azure's pricing page had not been updated yet (Microsoft Learn's list as of 2026-09-29). The full walkthrough, including real capacity figures, is in the Microsoft Foundry section.

Q. What does the Azure route cost?

A. Per Microsoft's Foundry blog (updated 2026-09-25): Direct from Azure is input $0.30 / output $1.20 / cached input $0.006, and Fireworks on Foundry is input $0.37 / output $1.50 / cached input $0.007 (USD per 1M tokens). Direct from Azure matches DeepSeek's peak rates, with no off-peak discount. Because Azure's pricing page has no V4.1-Flash row yet, confirm in the portal before committing.

Q. Does my code work as-is if I move to Azure?

A. Mostly, with two fixes. ① DeepSeek's own thinking parameter returns 400 on Azure; only reasoning_effort is accepted (measured 2026-09-28). ② The effort scale differs — the Foundry blog describes a continuous 1–100 setting, while the first-party API uses low/high/max. The endpoint is OpenAI-compatible (/openai/v1), so apart from those two points it is largely a drop-in swap.

Sources

• DeepSeek API Docs — Models & Pricing (pricing, specs, notes; retrieved 2026-09-11)
• DeepSeek API Docs — Change Log (2026-09-10, DeepSeek-V4.1-Flash Release) (release details, benchmarks)
• DeepSeek official news (news260910)
• Microsoft Foundry — model catalog (DeepSeek-V4.1-Flash) · Foundry blog (updated 2026-09-25) · Foundry Models quotas and limits · az cognitiveservices account deployment · Data, privacy, and security (Azure section; retrieved 2026-09-29)
• Related: pricing (DeepSeek) · token calculator · V4 Pro continuation · multimodal guide · release timeline