DeepSeek V4 Pro 对比 GPT-5.6：价格与规格实测

什么时候选哪个

两者都是文本输入的推理模型，上下文都在百万级（deepseek-v4-pro 为 1000000，gpt-5.6 为 1050000），思考可选，所以差别在价格、输出长度和模态。预算内做长文本生成选 deepseek-v4-pro：输入约便宜 3.8 倍，输出约便宜 7.6 倍，缓存读取约便宜 3.8 倍，并允许最多 393216 输出 token，对 128000。需要图像输入（deepseek-v4-pro 不接受）或者想要更新的 2026-07-09 代际和 2026-02 知识截止时，选 gpt-5.6。

Benchmark 成绩

领先高于同类均值无更高分DeepSeek V4 Pro05 ~~/ 10~~0 ~~/ 10~~GPT-5.6594 ~~/ 121~~28 ~~/ 121~~

5 项两边都有成绩。

DeepSeek V4 Pro GPT-5.6 其他参测模型同类均值 ★ 没有模型得分更高

SWE-Bench Pro

59%

64.6%

BioMysteryBench hard

N/A

44.7%

OSWorld-Verified

N/A

83%

Cybergym

N/A

84.5%

HealthBench Professional

N/A

60.5%

Finance Agent v2

N/A

53.8%

Harvey Lab-AA

N/A

87.2%

Humanity's Last Exam no tools

37.7%

44.5%

BrowseComp

N/A

90.4%

LVBench

N/A

82.1%

Humanity's Last Exam no tools

37.7%

44.5%

GPQA Diamond max

90.1%

没有模型得分更高 94.1%

IFBench max

77%

72.7%

GPQA Diamond

N/A

94.6%

Humanity's Last Exam with tools

N/A

58%

AA-LCR

N/A

73.7%

HLE-Verified

N/A

54.5%

FrontierMath Tier 4 (v2)

N/A

83%

ARC-AGI-3

N/A

7.78%

GraphWalks BFS 1M (F1)

N/A

没有模型得分更高 77.1%

CritPt

N/A

没有模型得分更高 32.3%

FrontierMath Tier 1-3 (v2)

N/A

没有模型得分更高 89%

GraphWalks BFS 256K (F1)

N/A

没有模型得分更高 90.7%

OpenAI MRCR v2 8-needle 512K-1M

N/A

73.8%

OpenAI MRCR v2 8-needle 256K-512K

N/A

91.5%

ARC-AGI-1

N/A

97.5%

ARC-AGI-2

N/A

92.5%

Apex Shortlist high

85.5%

N/A

Apex Shortlist max

90.2%

N/A

Chinese-SimpleQA high

77.7%

N/A

Chinese-SimpleQA max

84.4%

N/A

CorpusQA 1M high

56.5%

N/A

CorpusQA 1M max

62%

N/A

GPQA Diamond high

89.1%

N/A

HMMT Feb. 2026 high

94%

N/A

HMMT Feb. 2026 max

95.2%

N/A

Humanity's Last Exam no tools, high

34.5%

N/A

Humanity's Last Exam no tools, max

37.7%

N/A

Humanity's Last Exam with tools, high

44.7%

N/A

Humanity's Last Exam with tools, max

48.2%

N/A

IMOAnswerBench high

88%

N/A

IMOAnswerBench max

89.8%

N/A

MMLU-Pro high

87.1%

N/A

MMLU-Pro max

87.5%

N/A

MMMLU max

87.9%

N/A

MRCR 1M high

83.3%

N/A

MRCR 1M max

83.5%

N/A

SimpleQA-Verified high

46.2%

N/A

SimpleQA-Verified max

57.9%

N/A

Apex Shortlist non-think

9.2%

N/A

Chinese-SimpleQA non-think

75.8%

N/A

CorpusQA 1M non-think

35.6%

N/A

CritPt max

12.9%

N/A

Global PIQA max

90.5%

N/A

GPQA Diamond non-think

72.9%

N/A

HMMT Feb. 2026 non-think

31.7%

N/A

Humanity's Last Exam no tools, non-think

7.7%

N/A

IFEval max

91.9%

N/A

IMOAnswerBench non-think

35.3%

N/A

INCLUDE max

86.1%

N/A

MAXIFE max

88.9%

N/A

MMLU-Pro non-think

82.9%

N/A

MMLU-ProX max

83.9%

N/A

MMLU-Redux max

94.8%

N/A

MRCR 1M non-think

44.7%

N/A

MRCR v2 128K max

74.4%

N/A

NOVA-63 max

52.8%

N/A

PolyMATH max

72%

N/A

QwenWorldBench max

52.3%

N/A

SimpleQA-Verified non-think

45%

N/A

SuperGPQA max

69.9%

N/A

WMT24++ max

82.2%

N/A

SWE-Bench Pro

59%

64.6%

NL2Repo

35.5%

56.8%

SWE-Bench Pro max

55.4%

没有模型得分更高 64.6%

DeepSWE 1.1

N/A

72.7%

Terminal-Bench 2.1

N/A

88.8%

SWE-bench Multilingual

76.2%

N/A

Terminal-bench 4.0

N/A

37.3%

Terminal-bench 3.0

N/A

34.4%

SciCode

N/A

56.1%

FrontierCode 1.1 Main

N/A

47.5%

DeepSWE

N/A

没有模型得分更高 73%

ProgramBench Almost Solved

N/A

23%

MLS-Bench-Lite

N/A

46.2%

FrontierSWE

N/A

71.3%

Terminal Bench 2.0 Terminus 2

67.9%

N/A

Kimi Code Bench 2.0

N/A

64.8%

PostTrainBench

N/A

34.6%

ProgramBench

N/A

77.6%

FrontierCode 1.1 Extended

N/A

60.6%

SWE-Marathon

N/A

39%

SWE-Marathon v1.1

N/A

42.5%

CursorBench 3.2.0

N/A

67.2%

CursorBench 4.0

N/A

41.7%

Frontier-Bench v0.1

N/A

34.4%

Internal Database Migration Tasks

N/A

42.7%

Internal Research Debugging

N/A

没有模型得分更高 68.3%

KernelGen 1P

N/A

没有模型得分更高 61.1%

NanoGPT

N/A

9.69%

PostTrainBench Lite

N/A

50.3%

RSI Index

N/A

没有模型得分更高 57.9%

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

2919

N/A

LiveCodeBench high

89.8%

N/A

LiveCodeBench max

93.5%

N/A

SWE-bench Multilingual max

76.2%

N/A

SWE-Bench Pro high

54.4%

N/A

SWE-bench Verified high

79.4%

N/A

SWE-bench Verified max

80.6%

N/A

Terminal Bench 2.0 high

63.3%

N/A

Terminal Bench 2.0 max

67.9%

N/A

Terminal Bench 2.0 Terminus 2, max

67.9%

N/A

Codeforces (Elo) max · 3052-3206 据 DeepSeek 公布 · 2026-08-26

3206

N/A

LiveCodeBench non-think

56.8%

N/A

NL2Repo max

35.5%

N/A

QwenSVGBench Elo, max · 1506-1541 据 Alibaba (Qwen) 公布 · 2026-08-26

1506

N/A

QwenWebDev (Elo) Elo, max · 1570-1617 据 Alibaba (Qwen) 公布 · 2026-08-26

1570

N/A

SWE-bench Multilingual non-think

69.8%

N/A

SWE-bench Multilingual high

74.1%

N/A

SWE-Bench Pro non-think

52.1%

N/A

SWE-bench Verified non-think

73.6%

N/A

Terminal Bench 2.0 non-think

59.1%

N/A

DeepSWE 1.1 max 据 Alibaba (Qwen) 公布

没有其他模型的公开成绩

N/A

73%

MLS-Bench-Lite max 据 Alibaba (Qwen) 公布

没有其他模型的公开成绩

N/A

46.2%

Terminal-Bench 2.1 max 据 Alibaba (Qwen) 公布

没有其他模型的公开成绩

N/A

88.8%

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

价格

	DeepSeek V4 Pro	GPT-5.6	Δ
输入 / 1M token	$1.32	$5	0.26×
输出 / 1M token	$3.96	$30	0.13×
缓存读取 / 1M token	$0.132	$0.5	0.26×
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 76 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 Pro · $1.32 GPT-5.6 · $5

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

能力

	DeepSeek V4 Pro	GPT-5.6
工具调用	是	是
思考控制	可配置	可配置
结构化输出	是	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)	5-10m, up to 1h
最小缓存前缀	未公布	1024 个 token

规格

	DeepSeek V4 Pro	GPT-5.6
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	2026-07-09
知识截止日期	-	2026-02
上下文窗口	1M	1.1M
最大输出	393K	128K
思考参数	thinking.type reasoning_effort	reasoning.effort
可选值	thinking.type enabled disabled reasoning_effort high max (low and medium map to high, xhigh maps to max)	reasoning.effort none low medium high xhigh max
默认值	enabled, with reasoning_effort high some complex agent requests are automatically set to max	medium

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 Pro · GPT-5.6

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 Pro 通过 · 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.

输出 641 tok (+545 思考) 延迟 11.5 s

GPT-5.6 通过 · 3 sentences

Enabling reasoning makes the model generate many additional internal tokens before producing an answer, so each request consumes substantially more compute. Those reasoning tokens are billed even though they are not visible to users, and longer processing can also increase tool calls, retries, and output length. The bill doubled because our cost per request rose, not necessarily because usage volume increased; we can control it by limiting reasoning effort and routing simpler tasks to cheaper models.

输出 126 tok (+30 思考) 延迟 3.0 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 Pro 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 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. […]

输出 2418 tok (+2131 思考) 延迟 35.6 s

GPT-5.6 通过 · 8/8 cases

```python def longest_run(xs): if not xs: return 0 best = cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 else: best = max(best, cur) cur = 1 return max(best, cur) ``` The original code counted equal *transitions* rather than items, so a run of one had length `0`, a run of three had length `2`, and tracking from zero caused the off-by-one error. The corrected version starts each run at length `1`, resets to `1` when the value changes, and performs a final `max` after the loop so a run ending at the last element is included; an empty input still returns `0`.

输出 277 tok (+93 思考) 延迟 4.2 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 Pro 通过 · 5/5 fields, guidance null

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

输出 1204 tok (+1153 思考) 延迟 19.9 s

GPT-5.6 通过 · 5/5 fields, guidance "withheld pending the Q4 close"

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

输出 179 tok (+118 思考) 延迟 3.7 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 Pro 没有可供评分的回答 · no answer text within 16,384 tokens (all of it went to thinking)

模型没有返回回答文本，整个 token 预算都花在了隐藏的思考上。

输出 8193 tok (+8192 思考) 延迟 106.2 s

GPT-5.6 通过 · 120 words, 0 banned, 1 question

Today we are introducing Prompt Cache, a new API gateway feature that stores reusable prompt prefixes and serves them across supported model providers. By recognizing identical system instructions, tools, examples, and context blocks, Prompt Cache reduces repeated input processing, lowers latency, and helps control token costs without changing application code. Teams can configure retention windows, cache keys, provider routing, encryption, and regional storage policies from one dashboard or API. What happens when a preferred provider is unavailable? The gateway can route requests to another provider while reusing eligible cached content, preserving performance and consistency. Built-in metrics report hit rates, savings, latency, and provider usage, while audit logs support governance. Prompt Cache is available today in public preview for all customers.

输出 628 tok (+473 思考) 延迟 7.3 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-pro",
    # model="gpt-5.6",  # 取消注释此行，注释上一行
    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-pro",
  // model: "gpt-5.6",  // 取消注释此行，注释上一行
  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-pro",
    # "model": "gpt-5.6",  # 取消注释此行，注释上一行
    "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-pro",
		// Model: "gpt-5.6",  // 取消注释此行，注释上一行
		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-pro")
        // .model("gpt-5.6")  // 取消注释此行，注释上一行
        .addUserMessage("Summarize this diff")
        .reasoningEffort(ReasoningEffort.MEDIUM)
        .build());
System.out.println(resp.choices().get(0).message().content().orElse(""));

获取 API key →

常见问题

DeepSeek V4 Pro 和 GPT-5.6 哪个更便宜？

按「输入 / 1M token」算，DeepSeek V4 Pro 更便宜（$1.32 对 $5，相差 3.8×）。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 Pro 和 GPT-5.6 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 Pro 和 GPT-5.6 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

2026 最佳 Flash LLM API：GLM 性价比最高，Gemini 3.8 最强 LLM API 限流对比：13 家服务商，2 个 SDK 从不重试 429 AI API 计费单位：从 token 到 PTU-hour 的 17 种计量方式

DeepSeek V4 Pro vs GPT-5.6

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