DeepSeek V4 Flash (0731) vs DeepSeek V4 Pro：價格與規格實測

什麼情況選哪一個

這兩個 DeepSeek 同門在通常決定取捨的規格上一致——1000000 token 上下文、393216 最大輸出、文字進文字出，以及同樣的聊天、程式碼、推理和工具標記——所以真正的差別在價格和思考控制。deepseek-v4-flash-0731 在每一項費率上都更便宜，且幅度相近：輸入 $0.44 對 $1.32，輸出 $1.32 對 $3.96，快取讀取 $0.044 對 $0.132，各約 3 倍。想按次關閉思考時選 deepseek-v4-pro；單看價目表，包含快取前綴在內，Flash 全面勝出。

Benchmark 成績

高於平均DeepSeek V4 Flash (0731)10 ~~/ 28~~DeepSeek V4 Pro5 ~~/ 10~~

DeepSeek V4 Flash (0731) DeepSeek V4 Pro 其他有成績的模型其他模型平均 ★ 沒有模型分數更高

SWE-Bench Pro

56%

59%

Cybergym

76.7%

N/A

JobBench

41.3%

N/A

Humanity's Last Exam no tools

33.8%

37.7%

Agents' Last Exam

25.2%

N/A

Humanity's Last Exam no tools

33.8%

37.7%

GPQA Diamond

89.9%

N/A

Humanity's Last Exam with tools

51.5%

N/A

Humanity's Last Exam no tools, text-only

37.8%

N/A

IFBench

79.2%

N/A

MathArena Apex

58.6%

N/A

GPQA Diamond max

N/A

90.1%

Apex Shortlist high

N/A

85.5%

Apex Shortlist max

N/A

90.2%

Chinese-SimpleQA high

N/A

77.7%

Chinese-SimpleQA max

N/A

84.4%

CorpusQA 1M high

N/A

56.5%

CorpusQA 1M max

N/A

62%

GPQA Diamond high

N/A

89.1%

HMMT Feb. 2026 high

N/A

94%

HMMT Feb. 2026 max

N/A

95.2%

Humanity's Last Exam no tools, high

N/A

34.5%

Humanity's Last Exam no tools, max

N/A

37.7%

Humanity's Last Exam with tools, high

N/A

44.7%

Humanity's Last Exam with tools, max

N/A

48.2%

IFBench max

N/A

77%

IMOAnswerBench high

N/A

88%

IMOAnswerBench max

N/A

89.8%

MMLU-Pro high

N/A

87.1%

MMLU-Pro max

N/A

87.5%

MMMLU max

N/A

87.9%

MRCR 1M high

N/A

83.3%

MRCR 1M max

N/A

83.5%

SimpleQA-Verified high

N/A

46.2%

SimpleQA-Verified max

N/A

57.9%

Apex Shortlist non-think

N/A

9.2%

Chinese-SimpleQA non-think

N/A

75.8%

CorpusQA 1M non-think

N/A

35.6%

CritPt max

N/A

12.9%

Global PIQA max

N/A

90.5%

GPQA Diamond non-think

N/A

72.9%

HMMT Feb. 2026 non-think

N/A

31.7%

Humanity's Last Exam no tools, non-think

N/A

7.7%

IFEval max

N/A

91.9%

IMOAnswerBench non-think

N/A

35.3%

INCLUDE max

N/A

86.1%

MAXIFE max

N/A

88.9%

MMLU-Pro non-think

N/A

82.9%

MMLU-ProX max

N/A

83.9%

MMLU-Redux max

N/A

94.8%

MRCR 1M non-think

N/A

44.7%

MRCR v2 128K max

N/A

74.4%

NOVA-63 max

N/A

52.8%

PolyMATH max

N/A

72%

QwenWorldBench max

N/A

52.3%

SimpleQA-Verified non-think

N/A

45%

SuperGPQA max

N/A

69.9%

WMT24++ max

N/A

82.2%

供應商公布： Alibaba (Qwen) Anthropic ByteDance DeepSeek Google MiniMax Moonshot OpenAI Tencent Z.ai

定價

	DeepSeek V4 Flash (0731)	DeepSeek V4 Pro	Δ
輸入 / 1M tokens	$0.44	$1.32	0.33×
輸出 / 1M tokens	$1.32	$3.96	0.33×
快取讀取 / 1M tokens	$0.044	$0.132	0.33×
快取寫入	不額外計費	不額外計費	-

費率取自網站建置時的即時目錄；各模型頁面都列有最新的價目。

兩者的相對位置：每 1M tokens 的輸入價格，涵蓋同一計費單位下全部 76 個聊天模型（對數尺度）

DeepSeek V4 Flash (0731) · $0.44 DeepSeek V4 Pro · $1.32

$0.05 · Qwen3 VL Flash $30 · GPT-5.4 Pro

功能

	DeepSeek V4 Flash (0731)	DeepSeek V4 Pro
工具使用	是	是
思考控制	是，但供應商未公布調整參數	可設定
結構化輸出	是	是
提示詞快取	隱式（自動）	隱式（自動）
快取存活時間	no fixed TTL (evicted when unused)	no fixed TTL (evicted when unused)
最小快取前綴	未公布	未公布

規格

	DeepSeek V4 Flash (0731)	DeepSeek V4 Pro
輸入模態	文字	文字
輸出模態	文字	文字
發布日期	2026-07-31	2026-04-24
上下文視窗	1M	1M
最大輸出	393K	393K
思考參數	-	thinking.type reasoning_effort
可接受的值	-	thinking.type enabled disabled reasoning_effort high max (low and medium map to high, xhigh maps to max)
預設值	-	enabled, with reasoning_effort high some complex agent requests are automatically set to max

規格摘錄自各供應商的文件；供應商沒有公布的項目就直接略過，不自行推測。完整來源： DeepSeek V4 Flash (0731) · DeepSeek V4 Pro

同一段提示詞，兩個模型，經閘道實測

寫給 CFO 的推理修正有 bug 的函式依 schema 擷取資料依照限制條件撰寫

提示詞 Explain to a CFO, in exactly three sentences, why our LLM bill doubled after we enabled reasoning. 檢查恰好 3 句

DeepSeek V4 Flash (0731) 通過 · 3 sentences

Enabling reasoning mode prompts the model to generate extensive internal chain-of-thought tokens before answering, often multiplying total token output per query by several times. This directly inflates costs because we are billed per token, and these reasoning tokens are processed at a premium compute tier that carries a higher price per million tokens. However, this spend buys measurably higher accuracy and fewer failed calls, so we are tracking it against specific error-reduction KPIs to ensure the ROI justifies the increase.

輸出 365 tok (+266 思考) 延遲 5.1 s

DeepSeek V4 Pro 通過 · 3 sentences

Enabling reasoning causes the model to generate a lengthy internal chain‑of‑thought before it delivers a visible answer, and we are billed for every one of those hidden reasoning tokens just as we are for the final output. In most interactions, the volume of these unseen tokens at least equals the visible response, effectively doubling the total tokens consumed per query. That doubled token consumption translates directly into a doubled per‑query cost, which is why our overall LLM bill doubled.

輸出 641 tok (+545 思考) 延遲 11.5 s

是否照指令做（剛好三句，數得出來）、語氣是否貼合對象（寫給 CFO 看），以及下方 token 計量顯示出的隱藏思考計費落差。

提示詞 This function is supposed to return the longest run of consecutive equal items, but callers report it is off by one on single-element inputs and misses runs that end at the last index. Fix it and explain the bug in one paragraph. 檢查修正後通過測試

DeepSeek V4 Flash (0731) 通過 · 8/8 cases

```python def longest_run(xs): if not xs: return 0 best = 1 cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 best = max(best, cur) else: cur = 1 return best ``` The original bug stems from initializing `cur` to 0 and only incrementing it when adjacent elements are equal, which counts the number of *equal transitions* rather than the actual run length—so every run of length L is reported as L−1, and a single‑element run (with no transitions) yields 0. Moreover, because `best` is updated only inside the equality branch, it never accounts for a run that ends at the last element if that run is longer than any previously seen (though the real issue remains the off‑by‑one). […]

輸出 1268 tok (+1016 思考) 延遲 14.0 s

DeepSeek V4 Pro 通過 · 8/8 cases

```python def longest_run(xs): if not xs: return 0 best = 1 cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 best = max(best, cur) else: cur = 1 return best ``` **Explanation:** The bug stems from initializing `cur = 0` and only incrementing it when two adjacent elements match, which counts the number of *additional* equal elements rather than the actual run length (which is `cur + 1`). As a result, the returned `best` is always one less than the true longest run—most obviously returning 0 for a single-element list instead of 1. […]

輸出 2418 tok (+2131 思考) 延遲 35.6 s

修正是否真的正確（跑得起來）、說明的資訊密度，以及在範圍明確的任務上的 token 效率。

提示詞 Extract a JSON object with fields {company, quarter, revenue_usd_m, growth_pct, guidance} from this text. Use null for anything not stated; add no other fields. 檢查有效的 JSON，完全符合 schema

DeepSeek V4 Flash (0731) 通過 · 5/5 fields, guidance null

{ "company": "Meridian Robotics", "quarter": "Q3", "revenue_usd_m": 412, "growth_pct": null, "guidance": null }

輸出 1119 tok (+1072 思考) 延遲 12.3 s

DeepSeek V4 Pro 通過 · 5/5 fields, guidance null

{ "company": "Meridian Robotics", "quarter": "September quarter", "revenue_usd_m": 412, "growth_pct": "eleven and change", "guidance": null }

輸出 1204 tok (+1153 思考) 延遲 19.9 s

是否遵守 schema（不自創欄位）、面對幻覺誘因的表現（題目明確不給 guidance），以及結構化輸出路徑的差異。

提示詞 Write a 120-word product announcement for an API gateway feature that caches prompts across providers. Forbidden words: "seamless", "unlock", "game-changer", "revolutionize", "empower". Exactly one sentence must be a question. 檢查 120 詞，0 個禁用詞

DeepSeek V4 Flash (0731) 未通過 · 141 words, 0 banned, 1 question

**Introducing Prompt Cache for API Gateway** We are pleased to announce Prompt Cache, a new feature that stores and reuses identical prompt prefixes across all connected LLM providers. By automatically recognizing repeated requests, this cache reduces token consumption and lowers latency without altering your application logic. Whether you use OpenAI, Anthropic, or open-source models, the cache operates transparently, ensuring consistent behavior while cutting costs. Developers can configure cache TTL and scoping rules via a simple API call, giving you precise control over data freshness and privacy. This feature also minimizes redundant upstream calls, freeing provider rate limits for more critical workloads. Start caching today through the dashboard or CLI, and watch your operational expenses drop significantly. […]

輸出 254 tok (+80 思考) 延遲 4.4 s

DeepSeek V4 Pro 沒有可評分的回答 · no answer text within 16,384 tokens (all of it went to thinking)

模型沒有回傳任何回答文字，token 額度全都用在隱藏的思考上了。

輸出 8193 tok (+8192 思考) 延遲 106.2 s

是否遵守限制（字數上限、禁用詞清單、只能有一句問句）、文字風格的特徵，以及長度控制。

改一行程式碼就能在兩者之間切換

下面每個頁籤都列了這兩個模型 ID，要改的只有醒目標示的那兩行。端點、金鑰和請求格式都不變。

from openai import OpenAI

client = OpenAI(
    base_url="https://synthorai.io/v1",
    api_key="sk-syn-...",
)

resp = client.chat.completions.create(
    model="deepseek-v4-flash-0731",
    # model="deepseek-v4-pro",  # 取消這一行的註解，並把上一行註解掉
    messages=[{"role": "user", "content": "Summarize this diff"}],
    reasoning_effort="medium",
)
print(resp.choices[0].message.content)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://synthorai.io/v1",
  apiKey: "sk-syn-...",
});

const resp = await client.chat.completions.create({
  model: "deepseek-v4-flash-0731",
  // model: "deepseek-v4-pro",  // 取消這一行的註解，並把上一行註解掉
  messages: [{ role: "user", content: "Summarize this diff" }],
  reasoning_effort: "medium",
});
console.log(resp.choices[0].message.content);

curl https://synthorai.io/v1/chat/completions \
  -H "Authorization: Bearer sk-syn-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash-0731",
    # "model": "deepseek-v4-pro",  # 取消這一行的註解，並把上一行註解掉
    "messages": [{"role": "user", "content": "Hello"}],
    "reasoning_effort": "medium"
  }'

package main

import (
	"context"
	"fmt"

	"github.com/openai/openai-go/v3"
	"github.com/openai/openai-go/v3/option"
)

func main() {
	client := openai.NewClient(
		option.WithBaseURL("https://synthorai.io/v1"),
		option.WithAPIKey("sk-syn-..."),
	)
	resp, _ := client.Chat.Completions.New(context.TODO(), openai.ChatCompletionNewParams{
		Model: "deepseek-v4-flash-0731",
		// Model: "deepseek-v4-pro",  // 取消這一行的註解，並把上一行註解掉
		Messages: []openai.ChatCompletionMessageParamUnion{
			openai.UserMessage("Summarize this diff"),
		},
		ReasoningEffort: openai.ReasoningEffortMedium,
	})
	fmt.Println(resp.Choices[0].Message.Content)
}

import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.chat.completions.*;
import com.openai.models.ReasoningEffort;

OpenAIClient client = OpenAIOkHttpClient.builder()
    .baseUrl("https://synthorai.io/v1")
    .apiKey("sk-syn-...")
    .build();

ChatCompletion resp = client.chat().completions().create(
    ChatCompletionCreateParams.builder()
        .model("deepseek-v4-flash-0731")
        // .model("deepseek-v4-pro")  // 取消這一行的註解，並把上一行註解掉
        .addUserMessage("Summarize this diff")
        .reasoningEffort(ReasoningEffort.MEDIUM)
        .build());
System.out.println(resp.choices().get(0).message().content().orElse(""));

取得 API 金鑰 →

常見問題

DeepSeek V4 Flash (0731) 和 DeepSeek V4 Pro 哪個比較便宜？

以「輸入 / 1M tokens」來看，DeepSeek V4 Flash (0731) 比較便宜（$0.44 對 $1.32，相差 3.0×）。其他項目的結果可能相反，完整價目請看上表；實際成本要看你的用量組合。

可以只串接一次，就對 DeepSeek V4 Flash (0731) 和 DeepSeek V4 Pro 做 A/B 測試嗎？

可以。兩個模型都走同一個 OpenAI 相容端點，用的也是同一把 API 金鑰，切換時只要改一行裡的模型名稱字串。你可以把一部分流量分別導到兩邊，再直接比較帳單。

DeepSeek V4 Flash (0731) 與 DeepSeek V4 Pro 支援提示詞快取嗎？

支援。兩者的快取讀取費率都低於輸入費率，所以前綴已經進快取的工作負載，實際成本會比官網價算出來的低。確切的快取讀取價格請見上方定價表。

Benchmark 成績

定價

功能

規格

同一段提示詞，兩個模型，經閘道實測

改一行程式碼就能在兩者之間切換

常見問題

相關比較