DeepSeek API 400 "Model Not Exist": deepseek-chat and deepseek-reasoner were retired

What the error looks like

{
  "error": {
    "message": "Model Not Exist",
    "type": "invalid_request_error",
    "param": null,
    "code": "invalid_request_error"
  }
}

The status is 400, not 404, so many SDKs report it as a generic BadRequestError and the model id is not always in the message. If a request that has not changed starts failing with this body, the model id is the first thing to check.

The same body comes back for any id DeepSeek does not serve, including typos. What is new since July is that two ids that used to work are on that list.

Which ids changed, and what they map to now

The thinking-mode detail matters. deepseek-chat never returned reasoning tokens. deepseek-flash does by default, and reasoning tokens are billed as output. A rename without switching thinking off makes short requests slower and more expensive than they were.

Fix 1: rename the id at DeepSeek

Change the model string and, for anything that used deepseek-chat, disable thinking:

from openai import OpenAI

client = OpenAI(api_key="sk-...", base_url="https://api.deepseek.com")

# deepseek-chat was the NON-thinking mode. deepseek-flash thinks by default, so switch it off
# or you will pay for reasoning tokens you did not get before.
resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Summarise this in one line: ..."}],
    extra_body={"thinking": {"type": "disabled"}},
)

Then search your dependencies. LangChain, LlamaIndex, LiteLLM, Dify and many smaller wrappers shipped deepseek-chat as a default at some point. The GitHub reports for this error are mostly libraries and tools whose defaults had not been updated, not hand-written calls.

Fix 2: keep the old ids and change the base URL

DeepSeek V4 is an open-weight model, so hosts other than DeepSeek run it. We accept the retired ids as aliases and keep their behaviour: deepseek-chat runs with thinking off and deepseek-reasoner with thinking on, both on DeepSeek V4 Flash (the 0731 release). deepseek-flash maps to V4.1 Flash, as at DeepSeek. Tool calling and tool_choice work on all of them, including tool_choice: "required", which the old reasoner refused.

from openai import OpenAI

client = OpenAI(api_key="your-inferenceapis-key", base_url="https://api.inferenceapis.com/v1")

# "deepseek-chat" and "deepseek-reasoner" keep working. deepseek-chat answers without thinking,
# deepseek-reasoner thinks first, exactly as before. Both run DeepSeek V4 Flash (0731) on US GPUs.
resp = client.chat.completions.create(
    model="deepseek-chat",
    messages=[{"role": "user", "content": "Summarise this in one line: ..."}],
)
print(resp.choices[0].message.content)

for m in deepseek-chat deepseek-reasoner deepseek-flash; do
  curl -s https://api.inferenceapis.com/v1/chat/completions \
    -H "Authorization: Bearer $INFERENCEAPIS_API_KEY" -H "Content-Type: application/json" \
    -d "{\"model\":\"$m\",\"messages\":[{\"role\":\"user\",\"content\":\"Say ok\"}],\"max_tokens\":20}" \
    | python3 -c "import sys,json; d=json.load(sys.stdin); print('$m', '->', d['choices'][0]['message']['content'], d['usage'].get('completion_tokens_details'))"
done

The ids are listed by GET /v1/models, so client-side validation that checks the model list passes too.

What it costs, side by side

DeepSeek's prices are from its pricing page on September 17, 2026, and change by time of day. Ours are flat. If you are price-driven and can accept China hosting, DeepSeek off-peak is cheaper. If location, retention or a flat price matters, that is the case for a US host.

Frequently asked questions

Did DeepSeek announce the retirement?

Not in any place we could find. The docs were updated to list the new ids and the old ones stopped working. We date it from the first GitHub reports, July 24 to 26, 2026. If you know of an official notice, tell us and we will link it.

Is deepseek-flash the same model as deepseek-chat was?

No. deepseek-chat last pointed at DeepSeek V4 Flash; deepseek-flash is DeepSeek V4.1 Flash. DeepSeek also serves requests for the legacy deepseek-v4-flash id from V4.1 Flash now. If you need the 0731 build of V4 Flash specifically, it is deepseek-ai/DeepSeek-V4-Flash here.

Does deepseek-chat on Inference APIs return reasoning_content?

No. Thinking is off for that alias unless you send reasoning_effort or reasoning: {enabled: true} yourself. deepseek-reasoner returns reasoning_content on the message, and as delta.reasoning_content when streaming.

What about deepseek-coder?

That id was merged into deepseek-chat in 2024 and has not existed at DeepSeek since. We do not alias it.

Where Inference APIs fits

The second fix is what we sell. deepseek-chat, deepseek-reasoner, deepseek-v4-flash and deepseek-flash are all accepted ids on our endpoint, so a base URL change is the whole migration. The models run on GPUs in the United States with zero retention, billed per token from a prepaid balance. We are not cheaper than DeepSeek's own off-peak price and we say so in the table above.

Related entries

• OpenCode Go 403 RegionError: "only available hosted in China and requires explicit opt in"
• 404 model_not_found: llama-3.3-70b-versatile (Groq)
• 404 model_not_found: "The model … does not exist or you do not have access to it"