DeepSeek V4 Pro (V4-Pro-0813): Specs, Pricing and API Guide

DeepSeek V4 Pro today means DeepSeek-V4-Pro-0813, the official release that DeepSeek rolled out on 13 August 2026 to replace the V4-Pro Preview it had launched in April. In the API it is still called deepseek-v4-pro, it sits beside the cheaper deepseek-flash, and since 16 August 2026 it has been billed at peak and off-peak rates. This page sets out what the model is as of 23 October 2026, drawing only on DeepSeek’s API documentation, Change Log, news pages and Hugging Face model card: what changed from the Preview, how the low, high and max reasoning-effort levels work, what DeepSeek reports on benchmarks, what it costs next to deepseek-flash, and how to call it.

What DeepSeek V4 Pro is as of October 2026

The Hugging Face model card describes DeepSeek-V4-Pro-0813 as “the official release of DeepSeek-V4-Pro, superseding the preview version”, with greatly enhanced agentic capabilities and performance improvements that DeepSeek says are most pronounced in real-world use. It is built on the DeepSeek-V4-Pro (Preview) model structure, with a DSpark speculative-decoding module attached. The 13 August 2026 Change Log entry calls it the GA release and says it was rolled out on the app, the web and the API, with no change to how the API is called: requests that name deepseek-v4-pro get the new version.

DeepSeek’s Models & Pricing page lists the model as follows on 23 October 2026.

On DeepSeek’s app and web chat, the 13 August news page says V4 Pro is reached through “Expert Mode”. For the whole family, including the history of the Flash models, see the DeepSeek V4 overview.

From the April Preview to V4-Pro-0813

The model name has stayed the same since April while the model behind it has changed. DeepSeek’s Change Log and news pages give this sequence:

• 24 April 2026 – V4 Preview. deepseek-v4-pro and deepseek-v4-flash arrived in the API. The legacy names deepseek-chat and deepseek-reasoner were pointed at deepseek-v4-flash’s non-thinking and thinking modes and were scheduled for retirement on 24 July 2026 at 15:59 UTC; they are no longer listed in DeepSeek’s documentation.
• 31 July 2026 – V4-Flash-0731. The official V4-Flash went into API public beta. The Change Log noted that the V4-Pro API was unchanged and that “the official release of DeepSeek-V4-Pro will follow soon”.
• 13 August 2026 – V4-Pro GA (V4-Pro-0813). Three thinking-effort levels (low, high, max), native support for the OpenAI Responses API, and peak and off-peak pricing that took effect at 16:00 UTC on 16 August 2026, with off-peak at half the peak rate.
• 10 September 2026 – V4.1-Flash. A new deepseek-flash model with native image understanding; V4-Flash and V4-Flash-Vision-Exp were retired and prices were reduced.
• 17 October 2026 – V4.1-Flash-1017. deepseek-flash was upgraded again, with API calls and pricing unchanged, and DeepSeek opened its Search and Fetch APIs.

Is deepseek-v4-pro still served by V4-Pro?

Yes, according to DeepSeek’s current pages, but the record is worth reading carefully. The 10 September news page announced that V4-Pro was being phased out and that, from 04:00 UTC on 14 September 2026, all deepseek-v4-pro requests would route to V4.1-Flash at V4.1-Flash rates until a V4.1-Pro launched. The Change Log entry for the same date says the opposite: “In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged”, and that DeepSeek will give further notice of any change. The Models & Pricing page on 23 October 2026 lists DeepSeek-V4-Pro-0813 as the model version behind deepseek-v4-pro, at its own prices. We read the Change Log and the pricing page as the current position, and the 10 September routing plan as superseded.

The Preview’s parameter counts are history

DeepSeek’s 24 April 2026 announcement and the Preview model card gave the April Preview of V4-Pro as a Mixture-of-Experts model with 1.6T total parameters and 49B active, and the Preview of V4-Flash as 284B total and 13B active. The V4-Pro-0813 card does not restate a parameter count; it says only that the release is built on the Preview’s model structure plus the DSpark module. We therefore quote 1.6T / 49B as the Preview’s figures and do not present them as a specification of V4-Pro-0813.

Reasoning effort and thinking mode

Thinking mode is on by default, with a default effort of high. In the OpenAI format it is switched with "thinking": {"type": "enabled"} or {"type": "disabled"}, and the depth is set with "reasoning_effort", which accepts low, high or max. DeepSeek’s guidance in the 13 August Change Log is to use low for simple tasks, high for daily agent tasks and max for more complex scenarios. The Thinking Mode guide maps other effort names onto these three: minimal becomes low, medium and xhigh become high, and ultra becomes max. In the Anthropic format the effort goes in output_config, and in the Responses API format in reasoning, where none turns thinking off.

The Thinking Mode guide sets these parameter rules that apply to deepseek-v4-pro (the top_p rule is specific to it):

• In thinking mode, temperature, presence_penalty and frequency_penalty have no effect (they do not raise an error).
• In thinking mode your top_p is ignored: it is fixed at 0.95 when the request carries tools, and at 1.0 otherwise. In non-thinking mode top_p takes effect.
• The reasoning comes back in reasoning_content, next to content. When a request carries tools, the reasoning_content of earlier turns must be passed back, or the API returns a 400 error.

The API’s maximum output is 384K tokens, and the model card recommends a maximum output length of 384K tokens for the high and max effort levels.

Benchmarks DeepSeek reports for V4-Pro-0813

The figures below are DeepSeek-reported and copied exactly from the V4-Pro-0813 model card on Hugging Face (13 August 2026), including the other models’ columns as DeepSeek printed them. The site has not run any of these evaluations. According to the card, the code-agent tasks were run with the minimal mode of DeepSeek Harness as the agent framework, at max reasoning effort with temperature = 1.0, top_p = 0.95.

† The card describes DSBench-FullStack as an internal full-stack development set and DSBench-Hard as an internal set of difficult coding-agent problems. A dash means the card prints no score.

Read as DeepSeek presents it, the table shows two things. First, V4-Pro-0813 is well ahead of the April Preview on every row; on DeepSWE, for example, the card gives 62.7 against 12.8. Second, it does not top the table: at least one of Kimi K3, Opus-4.8 or Fable-5 (w/ fallback) has a higher figure than V4-Pro-0813 on eight of the ten rows, and V4-Pro-0813 has the highest figure only on Cybergym and AutomationBench (Public). That fits the card’s own wording, “broadly competitive with the strongest proprietary models”.

The cheaper model also matters here. DeepSeek’s 10 September news page says results from multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime, and the 10 September Change Log lists the September V4.1-Flash build at 90.6 on Terminal-Bench 2.1 and 88.1 on CyberGym, against 87.9 and 83.3 on the V4-Pro-0813 card. The two pages do not say the evaluation set-ups were identical, and the September build has since been replaced by V4.1-Flash-1017, so treat this as DeepSeek’s own indication rather than a like-for-like result.

No image input on deepseek-v4-pro

The Models & Pricing page lists Vision as “Not supported” for deepseek-v4-pro. Images go to deepseek-flash instead: its Vision column is ticked, and DeepSeek’s Vision guide says it accepts images alongside text in JPEG, PNG, GIF and WebP, sent as base64, as an external URL or through the Files API, in both Chat Completions and the Responses API. An application that needs both can send image steps to deepseek-flash and text-only steps to deepseek-v4-pro; both use the same base URL and API key. See the deepseek-flash model page for details.

DeepSeek V4 Pro pricing (as of 23 October 2026)

Prices are in US dollars per 1M tokens, from DeepSeek’s Models & Pricing page. Off-peak rates are half the peak rates. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; every other hour is off-peak, including weekends and Chinese public holidays in full.

The ratios are the same at peak and off-peak, because both models halve off-peak: deepseek-v4-pro costs about 7.3 times deepseek-flash for cache-hit input ($0.044 ÷ $0.006), 4.4 times for cache-miss input ($1.32 ÷ $0.30) and 3.3 times for output ($3.96 ÷ $1.20). Its concurrency limit of 500 is a fifth of deepseek-flash’s 2,500. The limit is counted per account, whichever API key is used; requests above it get an HTTP 429 error, and DeepSeek’s Rate Limit page says a capacity-expansion request costs nothing extra.

In thinking mode the model returns its reasoning in reasoning_content as part of its output, and DeepSeek bills the total number of input and output tokens, so long reasoning at max effort adds to the output side of the bill. Two separate items appear on the same page: the Search and Fetch APIs cost $0.75 per 1,000 successful requests each, and server-side tool calls made through DeepSeek Harness, Claude Code or Codex are free until per-request billing starts at 00:00 Beijing time on 1 November 2026 (16:00 UTC on 31 October 2026).

Worked example

An illustrative workload of 1,000,000 calls, each with a 2,000-token cached prefix (cache hit), 200 uncached input tokens (cache miss) and 300 output tokens. That is 2,000M cache-hit tokens, 200M cache-miss tokens and 300M output tokens in total.

deepseek-v4-pro, peak
  Cache hit : 2,000M × $0.044 = $88.00
  Cache miss:   200M × $1.32  = $264.00
  Output    :   300M × $3.96  = $1,188.00
  Total                         $1,540.00

deepseek-v4-pro, off-peak
  Cache hit : 2,000M × $0.022 = $44.00
  Cache miss:   200M × $0.66  = $132.00
  Output    :   300M × $1.98  = $594.00
  Total                         $770.00

deepseek-flash, peak
  Cache hit : 2,000M × $0.006 = $12.00
  Cache miss:   200M × $0.30  = $60.00
  Output    :   300M × $1.20  = $360.00
  Total                         $432.00

deepseek-flash, off-peak
  Cache hit : 2,000M × $0.003 = $6.00
  Cache miss:   200M × $0.15  = $30.00
  Output    :   300M × $0.60  = $180.00
  Total                         $216.00

For this mix deepseek-v4-pro costs about 3.6 times deepseek-flash ($1,540.00 ÷ $432.00 at peak, $770.00 ÷ $216.00 off-peak), and moving the whole job off-peak halves either bill. A mix with a larger output share, for example from long thinking-mode reasoning, moves the gap towards the 3.3× output ratio; a mix dominated by cache hits moves it towards 7.3×. To model your own token mix and hours, use the DeepSeek pricing calculator; the full rate card, including Search and Fetch, is on our DeepSeek API pricing page.

Calling deepseek-v4-pro

The API accepts the OpenAI Chat Completions format at https://api.deepseek.com, the Anthropic format at https://api.deepseek.com/anthropic, and, since 13 August 2026, the OpenAI Responses API natively. This Python example uses the OpenAI SDK (pip3 install openai) and reads the key from the DEEPSEEK_API_KEY environment variable, following DeepSeek’s own quick-start sample:

import os
from openai import OpenAI

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

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain the difference between a process and a thread."},
    ],
    stream=False,
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)

message = response.choices[0].message
print("Reasoning:", message.reasoning_content)
print("Answer:", message.content)

The same request with curl:

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{
        "model": "deepseek-v4-pro",
        "messages": [
          {"role": "system", "content": "You are a helpful assistant."},
          {"role": "user", "content": "Hello!"}
        ],
        "thinking": {"type": "enabled"},
        "reasoning_effort": "high",
        "stream": false
      }'

Set "reasoning_effort": "max" for the hardest tasks or "low" for quick ones, or send "thinking": {"type": "disabled"} for a plain non-thinking reply. Switching an existing integration between the two models only means changing the model field. Older code that still sends deepseek-v4-flash is accepted, but DeepSeek serves it with DeepSeek-V4.1-Flash-1017 at the Flash price, not with V4-Pro. For a first project, see the DeepSeek API getting started tutorial; for SDK details, see DeepSeek OpenAI SDK compatibility.

Open weights and self-hosting

The repository and weights of DeepSeek-V4-Pro-0813 are released under the MIT licence on Hugging Face. The card notes that the release has no Jinja-format chat template; instead an encoding folder holds Python scripts that turn OpenAI-style messages into model input and parse the output, with reasoning_effort taking the same three levels. For serving, the card enables DSpark in vLLM with --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}' (its example targets a single 4×GB300 node), and in SGLang with --speculative-algorithm DSPARK and no separate draft-model path, because the target and draft weights come from the same checkpoint. For local deployment it recommends temperature = 1.0, with top_p = 0.95 for agentic work and 1.0 otherwise. Licensing background is in is DeepSeek open source.

V4 Pro or deepseek-flash?

• Cost and throughput. deepseek-flash is cheaper on every price line (3.3× to 7.3× by the ratios above) and allows five times the concurrency.
• Images. Only deepseek-flash accepts image input.
• DeepSeek’s own position. Its 10 September news page put V4.1-Flash ahead of V4-Pro, while the Change Log kept V4-Pro in service “in response to user demand”. Teams whose workflows are built around V4-Pro can keep using it at today’s rates; DeepSeek has said it will announce any change.
• Same interface. Both models share the base URLs, the 1M context, the 384K maximum output, thinking and non-thinking modes, and the low / high / max effort levels, so comparing them on your own tasks takes one changed field.

For agentic coding set-ups, see DeepSeek for coding; for how DeepSeek compares with closed models, see DeepSeek vs Claude.

Frequently asked questions