DeepSeek API Migration: Move to V4.1 Flash

If your application still sends deepseek-v4-flash, a successful request no longer means it is using the original V4 Flash model. DeepSeek's current documentation routes that compatibility name to V4.1 Flash. New integrations should request deepseek-flash explicitly.

This guide updates the official API migration path as of September 23, 2026. It is for developers maintaining a script, server, or agent integration. Moving a conversation between web chat products is a separate task; copying chat history does not change the API configuration in your application.

Check what each identifier means now

The official change log and model table establish the current mapping:

Identifier in your configuration	Current documented status	Suggested action
`deepseek-flash`	V4.1 Flash	Use for new Flash integrations
`deepseek-v4-flash`	Compatibility alias for V4.1 Flash	Replace in maintained examples and configuration
`deepseek-v4-flash-vision-exp`	Compatibility alias for V4.1 Flash	Move image requests to `deepseek-flash`
`deepseek-v4-pro`	V4 Pro remains available	Evaluate before changing an established workflow
`deepseek-chat` / `deepseek-reasoner`	Officially scheduled for discontinuation on July 24, 2026	Migrate instead of depending on continued acceptance

The old chat and reasoner names should not be treated as the long-term way to choose non-thinking and thinking behavior. Make the model choice and thinking setting explicit in maintained code. Also inspect scheduled jobs and background workers: changing the interactive application does not update a model name copied into another service.

Separate the provider from the model

Write down the endpoint, requested model name, request format, and response fields your application consumes. A third-party gateway may use different aliases or prices. The mappings above describe DeepSeek's official service, not every service that offers a DeepSeek model.

For example, an environment variable named MODEL=deepseek-v4-flash tells you only one part of the story. You also need the base URL and the code that translates application messages into API requests. An adapter can drop fields or impose a smaller output limit even when the underlying model supports more.

Search configuration files, deployment variables, and prompt evaluation scripts before editing. Keep the provider unchanged for the first migration check. Changing providers, protocols, prompts, and models at once makes a failed result much harder to diagnose.

Start using deepseek-v4

Make one small request with an explicit mode

The following Python example calls the official Chat Completions endpoint using an OpenAI-compatible client. Install the openai package in your test environment and provide DEEPSEEK_API_KEY through your normal secret-management mechanism. Keep the key out of source files, browser code, and request logs.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["DEEPSEEK_API_KEY"],
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{
        "role": "user",
        "content": "Extract the order ID from: Order A-104 is ready. Return only the ID.",
    }],
    max_tokens=128,
    extra_body={"thinking": {"type": "disabled"}},
)

print(response.choices[0].message.content)
print(response.usage)

The task has an obvious expected answer, A-104. It checks authentication, routing, response parsing, and usage reporting without a complex prompt. It is a documentation-based example, not a recorded live API test. An unexpected answer or error is a reason to inspect the request before adding your full application context.

A 128-token cap is appropriate for this tiny non-thinking task. It is not a recommended budget for reasoning or agent work. When restoring your real workload, choose a budget that allows the requested output and check the completion's finish reason for truncation.

Thinking settings can change behavior without an error

The Thinking Mode documentation says thinking is enabled by default with high effort. In Chat Completions, thinking is supplied through extra_body when using the Python client. Supported effort levels are low, high, and max.

For a thinking-mode trial, replace the relevant arguments with:

reasoning_effort="high",
extra_body={"thinking": {"type": "enabled"}},

Do not assume a successful HTTP response means every sampling setting was applied. The documentation says temperature has no effect in thinking mode, even though it may be accepted for compatibility. If an older test suite relied on temperature to control variation, document that difference before comparing results.

Tool-calling history needs particular care. For Chat Completions requests carrying tools, the guide requires previous reasoning_content to be passed back, including turns without a tool call. An adapter that preserves only the assistant's final text can therefore break a later turn. Inspect the complete message handling in a multi-turn test, rather than validating only the first response.

Build the regression set around failure cases

Start with examples whose expected results you already know. Include an extraction with a missing field, a response that must fit a schema, a long input with the answer near the end, and a tool workflow that continues after a tool result. Add an image case only if your application uses images.

For each task, check whether the output is usable before measuring latency. A JSON-looking answer still needs parsing and schema validation. A code change still needs the relevant tests. A cited value still needs to exist in the supplied material. Save failure categories so you can distinguish an output-format regression from a provider error.

Measure completed work, including retries. Record input, output, and cache usage separately. A price reduction does not guarantee a cheaper task if the new behavior produces longer answers or causes your application to retry. Use identical prompts and explicit settings for the comparison; leave prompt optimization for a later pass.

Pro availability and rollback need separate decisions

The September 10 announcement described a plan to route Pro requests to Flash. The current change log says Pro service continues after September 14 with its billing method unchanged. Do not migrate a working Pro integration solely because you read the older retirement plan.

For Flash, reverting the configuration to deepseek-v4-flash does not restore the retired model: that identifier is an alias for the new model. A rollback plan must name an actually available service and account for its capabilities. Pro is not a drop-in fallback for an image request; the current official model table does not list vision support for Pro.

Roll out the change to a small portion of traffic first, with a clear condition for stopping. Keep the evaluation results and provider settings with the release record. Once the new route passes your application's checks, update copied examples and internal documentation so the old identifier does not return during the next integration.