DeepSeek Chat Model: What `deepseek-chat` Means Now

If you’re choosing a DeepSeek model for an API-powered chat product, don’t start a new integration with deepseek-chat. DeepSeek said that legacy model name would become inaccessible after July 24, 2026, at 15:59 UTC. Its current model and pricing documentation lists deepseek-flash and deepseek-v4-pro; it identifies Flash as DeepSeek-V4.1-Flash-1017. Both models are listed with a 1-million-token context window and a maximum output of 384,000 tokens. (DeepSeek model and pricing documentation; V4 announcement)

For a new text-chat integration, evaluate deepseek-flash first. DeepSeek documents image input for Flash, while its model comparison marks vision as unsupported for V4 Pro. The same comparison lists tool calls and both thinking and non-thinking modes for each model. Which model and mode to ship depends on your own quality, latency, and cost tests. (DeepSeek model and pricing documentation; Vision guide)

What does deepseek-chat refer to?

deepseek-chat was a legacy API model name. DeepSeek’s V4 announcement said it routed to V4-Flash in non-thinking mode during the transition and would be retired after July 24, 2026, at 15:59 UTC. That deadline has passed. For new requests, use a currently listed model ID—deepseek-flash or deepseek-v4-pro—rather than relying on the old name. (DeepSeek V4 announcement; current model documentation)

If you maintain an older app, search its configuration and requests for deepseek-chat and deepseek-reasoner, replace those identifiers with a currently documented model ID, and test the updated integration. DeepSeek’s documentation names deepseek-flash for Flash. (DeepSeek model documentation)

Which current DeepSeek model fits chat?

Model ID	Consider it when…	Documented details
deepseek-flash	You want to evaluate Flash for chat, or your chat product needs image input.	Version listed: DeepSeek-V4.1-Flash-1017. Supports image input, tool calls, and thinking and non-thinking modes.
deepseek-v4-pro	Your evaluation shows it better meets your needs and image input is not required.	Version listed: DeepSeek-V4-Pro-0813. Supports tool calls and thinking and non-thinking modes; the model comparison lists vision as unsupported.

DeepSeek lists a 1M-token context length and a maximum output of 384K tokens for the models. These are limits, not a recommendation to send very long prompts; large requests can increase usage costs and may not improve your results. Check the documentation again when planning production capacity because specifications and prices can change. (DeepSeek model and pricing documentation)

Thinking mode is enabled by default. DeepSeek documents a control for disabling it, as well as reasoning-effort settings. For ordinary conversational turns, compare the default with thinking disabled before choosing what to ship; test thinking enabled where additional reasoning may help. The documentation does not establish that one setting is best for every chat workload. (DeepSeek Thinking Mode guide)

If your chat product accepts images, Flash is the documented choice between these two models: DeepSeek’s vision guide says deepseek-flash accepts image input, including JPEG, PNG, GIF, and WebP. DeepSeek’s model comparison marks vision as unsupported for V4 Pro. (DeepSeek Vision guide; model and pricing documentation)

DeepSeek API pricing for Flash and V4 Pro

DeepSeek lists prices in US dollars per 1 million tokens. Input prices differ for cache hits and cache misses, and input and output prices also vary by peak and off-peak periods. The table reproduces the listed rates. (DeepSeek model and pricing documentation)

Model	Input, cache hit (off-peak / peak)	Input, cache miss (off-peak / peak)	Output (off-peak / peak)
deepseek-flash	$0.003 / $0.006	$0.15 / $0.30	$0.60 / $1.20
deepseek-v4-pro	$0.022 / $0.044	$0.66 / $1.32	$1.98 / $3.96

DeepSeek defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours—including weekends and Chinese public holidays—are off-peak. Its documentation says off-peak rates are half the peak rates and that prices may change. It does not establish here a separate rule for assigning a request to a rate period, so check the pricing documentation when estimating a production budget. (DeepSeek pricing documentation)

Worked example: Suppose a request uses 100,000 uncached input tokens and generates 20,000 output tokens. At Flash’s off-peak rates, the illustrative model charge is $0.027: (0.1 × $0.15) + (0.02 × $0.60). At peak rates, it is $0.054. The same token counts on V4 Pro cost $0.1056 off-peak or $0.2112 at peak. These estimates assume all input tokens are cache misses; actual token usage, caching, and prices affect the bill. (DeepSeek pricing documentation)

How to evaluate a model for production chat

Treat model choice as a workload test, not a family-name decision. A practical evaluation:

1. Collect representative prompts from your intended product, including routine turns, ambiguous questions, edge cases, and any image or tool-use requests you expect to support.
2. Run the same prompts on deepseek-flash and, if useful, deepseek-v4-pro. For Flash, compare thinking enabled and disabled where relevant.
3. Score the outputs against your own acceptance criteria—for example, factual correctness, instruction-following, useful refusals, and valid structured output if your app requires it.
4. Measure production trade-offs: record latency, input and output token counts, estimated cost using the applicable rates, and failure or retry rates.
5. Choose conditionally: use the least costly model and mode that meets your quality and operational requirements. If Flash fails important cases, compare Pro on those cases; if neither is acceptable, reconsider the model choice or product workflow.

This is an evaluation approach, not a claim that either model will achieve a particular quality or latency. Published specifications and prices do not replace tests on your prompts, application settings, and expected traffic.

Minimal API check

DeepSeek documents an OpenAI-compatible Chat Completions endpoint at https://api.deepseek.com/chat/completions, with https://api.deepseek.com as the SDK base URL. This Python example uses Flash with thinking disabled, using the documented extra_body control. It assumes the OpenAI Python SDK is installed and DEEPSEEK_API_KEY is set in the environment. (DeepSeek API quick start; Thinking Mode guide)

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": "system", "content": "You are a concise support assistant."},
        {"role": "user", "content": "How do I reset my password?"},
    ],
    extra_body={"thinking": {"type": "disabled"}},
)

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

This verifies a basic text-chat request, not production readiness. Before launch, test error handling, usage and billing, expected concurrency, and the conversation history your app will send. DeepSeek describes its chat API as stateless, so clients must include the prior conversation context needed for each request. (DeepSeek Chat Completions API)

Bottom line

For a new DeepSeek API chat integration, use deepseek-flash as the initial candidate, not the retired deepseek-chat name. Compare thinking settings on representative requests, and test deepseek-v4-pro if Flash does not meet your requirements. Verify current rates and API details before launch, especially if caching or peak-hour usage affects your cost estimate. (DeepSeek V4 announcement; current model and pricing documentation)

Sources

Models & Pricing | DeepSeek API Docs

DeepSeek V4 Preview Release | DeepSeek API Docs

Vision | DeepSeek API Docs

Thinking Mode | DeepSeek API Docs

Chat Completions API | DeepSeek API Docs

Your First API Call | DeepSeek API Docs