DeepSeek V4 Flash 对比 DeepSeek V4.1 Flash：3 倍价差，上下文同为 1M

什么时候选哪个

两者均支持 1,000,000 token 的上下文和 393,216 token 的最大输出，且均支持工具调用和推理，因此升级与否取决于输入类型和价格：deepseek-v4.1-flash 增加了图像输入，输入价格为 $0.3（约 2.2x），输出价格为 $1.2（约 4.4x），而仅支持文本的 deepseek-v4-flash 的输入和输出价格分别为 $0.138 和 $0.275，不过前者的缓存读取更便宜，为 $0.006，后者为 $0.0276。对于关注更低输入和输出费率以及可以选择关闭思考功能的大发送量文本对话和代码任务，请继续使用 deepseek-v4-flash；当需要发送图像时，请迁移到 deepseek-v4.1-flash。

Benchmark 成绩

高于同类均值无更高分DeepSeek V4 Flash仅 2 项可比DeepSeek V4.1 Flash14 ~~/ 19~~4 ~~/ 19~~

DeepSeek V4 Flash DeepSeek V4.1 Flash 其他参测模型同类均值 ★ 没有模型得分更高

DeepSWE 1.1

N/A

没有模型得分更高 74.2%

Cybergym

N/A

没有模型得分更高 88.1%

GPQA Diamond

N/A

90.9%

Agents' Last Exam

N/A

31.8%

Chartography with tools

N/A

78.9%

GPQA Diamond

N/A

90.9%

Humanity's Last Exam no tools

N/A

36.8%

Humanity's Last Exam with tools

N/A

63.9%

Humanity's Last Exam no tools, text-only

N/A

39.1%

MathArena Apex

N/A

65.6%

GPQA Diamond max

88.1%

N/A

Apex Shortlist high

72.1%

N/A

Apex Shortlist max

85.7%

N/A

Chinese-SimpleQA high

73.2%

N/A

Chinese-SimpleQA max

78.9%

N/A

CorpusQA 1M high

59.3%

N/A

CorpusQA 1M max

60.5%

N/A

GPQA Diamond high

87.4%

N/A

HMMT Feb. 2026 high

91.9%

N/A

HMMT Feb. 2026 max

94.8%

N/A

Humanity's Last Exam no tools, high

29.4%

N/A

Humanity's Last Exam no tools, max

34.8%

N/A

Humanity's Last Exam with tools, high

40.3%

N/A

Humanity's Last Exam with tools, max

45.1%

N/A

IMOAnswerBench high

85.1%

N/A

IMOAnswerBench max

88.4%

N/A

MMLU-Pro high

86.4%

N/A

MMLU-Pro max

86.2%

N/A

MRCR 1M high

76.9%

N/A

MRCR 1M max

78.7%

N/A

SimpleQA-Verified high

28.9%

N/A

SimpleQA-Verified max

34.1%

N/A

Apex Shortlist non-think

9.3%

N/A

Chinese-SimpleQA non-think

71.5%

N/A

CorpusQA 1M non-think

15.5%

N/A

GPQA Diamond non-think

71.2%

N/A

HMMT Feb. 2026 non-think

40.8%

N/A

Humanity's Last Exam no tools, non-think

8.1%

N/A

IMOAnswerBench non-think

41.9%

N/A

MMLU-Pro non-think

83%

N/A

MRCR 1M non-think

37.5%

N/A

SimpleQA-Verified non-think

23.1%

N/A

DeepSWE 1.1

N/A

没有模型得分更高 74.2%

NL2Repo

N/A

64%

Terminal-Bench 2.1

N/A

没有模型得分更高 90.6%

Terminal-bench 4.0

N/A

31.2%

Terminal-bench 3.0

N/A

30%

ProgramBench Almost Solved

N/A

20.3%

SWE-Bench Pro max

52.6%

N/A

Codeforces (Elo) high · 2816-3052 据 DeepSeek 公布 · 2026-08-26

2816

N/A

Codeforces (Elo) max · 3289-3471 据 DeepSeek 公布 · 2026-09-14

N/A

3471

LiveCodeBench high

88.4%

N/A

LiveCodeBench max

91.6%

N/A

SWE-bench Multilingual max

73.3%

N/A

SWE-Bench Pro high

52.3%

N/A

SWE-bench Verified high

78.6%

N/A

SWE-bench Verified max

79%

N/A

Terminal Bench 2.0 high

56.6%

N/A

Terminal Bench 2.0 max

56.9%

N/A

LiveCodeBench non-think

55.2%

N/A

SWE-bench Multilingual non-think

69.7%

N/A

SWE-bench Multilingual high

70.2%

N/A

SWE-Bench Pro non-think

49.1%

N/A

SWE-bench Verified non-think

73.7%

N/A

Terminal Bench 2.0 non-think

49.1%

N/A

DeepSWE 1.1 Claude Code 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

69.8%

DeepSWE 1.1 Codex 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

65.6%

DeepSWE 1.1 DSH Minimal 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

72.6%

DeepSWE 1.1 DSH PTC 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

67.6%

DeepSWE 1.1 DSH Standard 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

70.5%

DeepSWE 1.1 mini-SWE 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

74.2%

DeepSWE 1.1 OpenCode 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

65.5%

DeepSWE 1.1 Pi 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

66.2%

Terminal-Bench 2.1 Claude Code 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

88%

Terminal-Bench 2.1 Codex 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

84.1%

Terminal-Bench 2.1 DSH Minimal 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

90.6%

Terminal-Bench 2.1 DSH PTC 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

85.8%

Terminal-Bench 2.1 DSH Standard 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

85.8%

Terminal-Bench 2.1 mini-SWE 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

90.3%

Terminal-Bench 2.1 OpenCode 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

85%

Terminal-Bench 2.1 Pi 据 DeepSeek 公布

没有其他模型的公开成绩

N/A

86.1%

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

价格

	DeepSeek V4 Flash	DeepSeek V4.1 Flash	Δ
输入 / 1M token	$0.138	$0.3	0.46×
输出 / 1M token	$0.275	$1.2	0.23×
缓存读取 / 1M token	$0.0276	$0.006	4.6×
缓存写入	不单独收费	不单独收费	-

价格取自构建时的实时目录，最新价格见各模型页面。

两个模型所处的位置：全部 80 个按同一单位计费的聊天模型的每 1M token 输入价分布（对数刻度）

DeepSeek V4 Flash · $0.138 DeepSeek V4.1 Flash · $0.3

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

能力

	DeepSeek V4 Flash	DeepSeek V4.1 Flash
工具调用	是	是
思考控制	可配置	是，但供应商未公布调节参数
结构化输出	是	-
提示词缓存	隐式（自动）	隐式（自动）
缓存有效期	no fixed TTL (evicted when unused)	no fixed TTL (evicted when unused)
最小缓存前缀	未公布	未公布

规格

	DeepSeek V4 Flash	DeepSeek V4.1 Flash
输入模态	文本	文本图像
输出模态	文本	文本
发布日期	2026-04-24	2026-09-10
上下文窗口	1M	1M
最大输出	393K	393K
思考参数	thinking.type reasoning_effort	-
可选值	thinking.type enabled disabled reasoning_effort high max (low and medium map to high, xhigh maps to max)	-
默认值	enabled, with reasoning_effort high some complex agent requests are automatically set to max	-

规格照录自各供应商的文档；供应商没有公布的项目，对应的行直接省略，不做推断。完整来源： DeepSeek V4 Flash · DeepSeek V4.1 Flash

同一条提示词，两个模型，经网关实测

面向 CFO 的推理修复含 bug 的函数按 schema 提取按约束条件写作

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

DeepSeek V4 Flash 通过 · 3 sentences

When reasoning is enabled, the model generates additional "chain-of-thought" tokens internally before producing the final answer, effectively doubling the total token count per request. Since our billing is based on tokens processed, this directly doubled your usage. Additionally, reasoning often requires more expensive model tiers or higher compute resources, further amplifying the cost increase.

输出 154 tok (+85 思考) 延迟 3.1 s

DeepSeek V4.1 Flash 通过 · 3 sentences

Enabling reasoning changed the unit economics: the model now spends extra billable tokens internally to “think” before producing the answer, and those reasoning tokens are charged like output tokens even though users never see them. That added token generation—plus longer contexts and occasional retries on harder prompts—roughly doubled our cost per request. We can control this by capping reasoning depth, routing only complex queries to reasoning models, and caching common prompts to bring the bill back down.

输出 423 tok (+327 思考) 延迟 14.6 s

指令遵循（恰好三句，数得出来）、受众适配（对 CFO 说话的口吻），以及下方 token 计数暴露出的隐藏思考计费差额。

提示词 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. 检查修复通过测试

DeepSeek V4 Flash 通过 · 8/8 cases

The bug is that the function counts the number of consecutive equal **pairs** (i.e., `cur` increments for each `i` where `xs[i] == xs[i-1]`), but it should return the **length** of the longest run. For a run of length `L`, the number of pairs is `L-1`, so the result is always one less than the correct value. This causes the function to return 0 for a single‑element list (expected 1) and to undercount runs that end at the last index (e.g., `[1,1]` returns 1 instead of 2). Additionally, runs of length 1 that are not part of any pair (such as a single element at the end of a list) are never considered because `best` is only updated when a pair is found. […]

输出 2494 tok (+2227 思考) 延迟 19.4 s

DeepSeek V4.1 Flash 通过 · 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 else: cur = 1 best = max(best, cur) return best ``` The bug is that `cur` is counting equal adjacent pairs, not elements in the current run: every time `xs[i] == xs[i - 1]`, the run length increases by one element, but the first element of the run was never counted, so all runs are undercounted by one. Also, `best` starts at `0`, so single-element inputs return `0` instead of `1`, and a one-element run ending at the last index is missed entirely.

输出 918 tok (+733 思考) 延迟 11.4 s

修复是否真的正确（能运行）、解释的信息密度，以及在一个范围明确的任务上的 token 效率。

提示词 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. 检查合法 JSON，严格符合 schema

DeepSeek V4 Flash 通过 · 5/5 fields, guidance null

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

输出 1101 tok (+1054 思考) 延迟 10.2 s

DeepSeek V4.1 Flash 通过 · 5/5 fields, guidance "withheld pending the Q4 close"

{"company":"Meridian Robotics","quarter":"September quarter","revenue_usd_m":412,"growth_pct":11.0,"guidance":"withheld pending the Q4 close"}

输出 1707 tok (+1667 思考) 延迟 15.1 s

是否严守 schema（不臆造字段）、能否顶住幻觉压力（原文明说暂不给出 guidance），以及结构化输出路径的差异。

提示词 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. 检查 120 词，0 个禁用词

DeepSeek V4 Flash 通过 · 119 words, 0 banned, 1 question

We're thrilled to introduce Prompt Cache, a new feature for our API Gateway that stores prompt responses from any provider. When a user submits a prompt you've already processed, the system returns the cached result instantly—eliminating redundant API calls and reducing latency. This works across OpenAI, Anthropic, Google, and more, with automatic cache invalidation based on your rules. What does this mean for your budget? Fewer API calls directly lower your monthly spend. Additionally, response times drop by up to 80% for cached prompts, improving user experience. Developers can configure cache duration per provider, set TTLs, and bypass cache when needed. The feature is available now in your gateway dashboard. Start saving time and money with Prompt Cache.

输出 2039 tok (+1887 思考) 延迟 15.6 s

DeepSeek V4.1 Flash 通过 · 116 words, 0 banned, 1 question

Today we're launching Cross-Provider Prompt Cache for our API gateway. It stores identical prompt requests and their responses across supported model providers, then serves cached results when a match is found. Teams can cut duplicate inference costs, reduce latency, and keep behavior consistent during provider failover. The cache works with configurable TTLs, per-route rules, and cache-key controls, so you decide what is reusable and what must stay fresh. Does your application send the same prompts to multiple providers? Now your gateway can answer many of those calls without another upstream request. Existing observability dashboards show hit rates, saved tokens, and estimated spend reduction. Enable it in the gateway console, set your policy, and start caching today.

输出 917 tok (+770 思考) 延迟 9.0 s

是否守住约束（字数预算、禁用词表、只能有一个问句）、文风特征，以及长度控制。

改一行代码就能在两个模型之间切换

下方每个标签页里都有两个模型 ID，高亮的那两行是唯一要改的地方。端点不变，API key 不变，请求结构也不变。

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",
    # model="deepseek-v4.1-flash",  # 取消注释此行，注释上一行
    messages=[{"role": "user", "content": "Summarize this diff"}],
)
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",
  // model: "deepseek-v4.1-flash",  // 取消注释此行，注释上一行
  messages: [{ role: "user", content: "Summarize this diff" }],
});
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",
    # "model": "deepseek-v4.1-flash",  # 取消注释此行，注释上一行
    "messages": [{"role": "user", "content": "Hello"}]
  }'

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",
		// Model: "deepseek-v4.1-flash",  // 取消注释此行，注释上一行
		Messages: []openai.ChatCompletionMessageParamUnion{
			openai.UserMessage("Summarize this diff"),
		},
	})
	fmt.Println(resp.Choices[0].Message.Content)
}

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

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")
        // .model("deepseek-v4.1-flash")  // 取消注释此行，注释上一行
        .addUserMessage("Summarize this diff")
        .build());
System.out.println(resp.choices().get(0).message().content().orElse(""));

获取 API key →

常见问题

DeepSeek V4 Flash 和 DeepSeek V4.1 Flash 哪个更便宜？

按「输入 / 1M token」算，DeepSeek V4 Flash 更便宜（$0.138 对 $0.3，相差 2.2×）。其他计费项的结论可能相反，完整价格见上方表格，实际成本取决于你的用量构成。

不用分别集成两次，就能对 DeepSeek V4 Flash 和 DeepSeek V4.1 Flash 做 A/B 测试吗？

可以。两个模型走同一个 OpenAI 兼容端点，用同一个 API key，切换时只要改一行里的模型名，所以可以给两个模型各分一部分流量，直接对比账单。

DeepSeek V4 Flash 和 DeepSeek V4.1 Flash 支持提示词缓存吗？

支持。两个模型的缓存读取价都低于各自的输入价，所以前缀能反复命中缓存的负载，实际成本会比按官网价估算的低。具体的缓存读取价见上方价格表。

我们的实测研究

OpenAI 提示词对象 11 月 30 日停用：保留 pmpt_ ID，补齐版本 2026 年 LLM API 价格变化：token 涨价，同等能力便宜 50 倍 DeepSeek V4.1 Flash 1017：关闭思考，完成 30/30 项小型编程任务

DeepSeek V4 Flash vs DeepSeek V4.1 Flash

Benchmark 成绩

价格

能力

规格

同一条提示词，两个模型，经网关实测

改一行代码就能在两个模型之间切换

常见问题

相关对比

我们的实测研究