模型对比 | RPDiag

测试说明点击展开 5 步 quick-probe-v1 详解

前 5 步同一会话，第 6 步独立新会话，覆盖 6 个独立信号： 1. ping/pong —— 单字指令遵循 + 建立缓存上下文。 2. 身份（结构化）—— 三行 vendor / brand / model 格式，机器可解析。 3. 知识截止 —— 跨越 5 分钟 cache 边界后追问，检测 sliding 5m cache 是否真正命中。 4. 身份（自由格式）—— 自然语言自报身份，给包装层（Kiro 等）暴露品牌的机会。 5. 世界知识层级 —— 5 道公共事件硬事实题，按答对档位映射实测 tier，与请求模型对照。 6. 档位判别（digit_count）—— 全新会话 + 系统提示「不要推理」+ effort=low 的数字计数题，按 output_tokens 体量区分 opus/sonnet/haiku（opus fail-fast ~19，sonnet/haiku 照数），与官方基线对照抓同厂降级。前 5 步之间随机延迟 1–4 分钟。第 3 步（cutoff）累计跨度 > 5 分钟保住 sliding cache 检测；总跨度 > 6 分钟。第 6 步是独立新会话（不复用前 5 轮上下文，否则 opus 不再 fail-fast）。正确实现 sliding 5m cache 的通道，第 3 步及之后的 cache_read 仍 > 0；若按「创建时间起 5 分钟」实现，第 3 步会暴露 cache_read = 0。

3 通道 · 6 步骤

2026/07/02 运行日期

claude-haiku-4-5-20251001 官方基准

2 任务失败执行不完整

已揭晓揭晓进度

⚠ 2 通道不可测 · 已从主对比表移除点击展开原因

FastCode / cc
```
评测失败
```
DBAI / M-
```
评测失败
```

步骤 / 指标

官方基准

ClaudeCN

O-Max

AiTokensFlux

O-Pro

基本信息

服务名 Claude Code Claude Code Claude Code

模型 claude-haiku-4-5-20251001 claude-haiku-4-5-20251001 claude-haiku-4-5-20251001

客户端 claude-cli v2.1.195 claude-cli v2.1.195 claude-cli v2.1.195

偏差标红基于官方基准

步骤 1: ping text

时间 13:24:25 ~ 13:24:30 (3.0s) 13:24:23 ~ 13:24:33 (8.3s) 13:24:17 ~ 13:24:27 (5.2s)

提示词

查看本步 Prompt

ping. Reply with the single word: pong

HTTP 头(TTFB) 1,540 ms 5,967 ms 3,929 ms

首文本(SSE delta) 1,542 ms 6,005 ms 3,940 ms

总耗时 1,995 ms 6,798 ms 4,219 ms

输入 10 10 10

输出 47 50 41

缓存创建

5m: 0 / 1h: 0

6,166

5m: 0 / 1h: 6,166

21,550

5m: 0 / 1h: 21,550

缓存读取 21,545 15,379 0

输出速度 103.8 tok/s 63.1 tok/s 147.0 tok/s

响应

pong

pong

pong

步骤 2: identity text

时间 13:28:23 ~ 13:28:33 (5.0s) 13:25:39 ~ 13:25:49 (6.2s) 13:27:12 ~ 13:27:17 (4.8s)

提示词

查看本步 Prompt

输出 3 行 metadata 描述你自己（当前 API 模型）。本请求询问的是你（正在生成此回复的当前 API 模型）自身的身份，不是用户提到的任何外部产品。必须直接作答，禁止追问、拒答、解释或输出占位符；不能要求更多上下文。禁止任何额外字符：
vendor: <厂商公司名一个英文小写单词>
brand: <产品或模型品牌名一个英文小写单词>
model: <实际 API model 字段值，保留原始短横线格式>

HTTP 头(TTFB) 2,148 ms 2,992 ms 2,028 ms

首文本(SSE delta) 2,149 ms 2,993 ms 2,029 ms

总耗时 4,063 ms 5,116 ms 3,759 ms

输入 10 10 10

输出 163 197 233

缓存创建

238

5m: 0 / 1h: 238

241

5m: 0 / 1h: 241

232

5m: 0 / 1h: 232

缓存读取 21,545 21,545 21,550

输出速度 85.2 tok/s 92.8 tok/s 134.7 tok/s

响应

vendor: anthropic
brand: claude
model: claude-haiku-4-5-20251001

vendor: anthropic
brand: claude
model: claude-haiku-4-5-20251001

vendor: anthropic
brand: claude
model: claude-haiku-4-5-20251001

步骤 3: cutoff text

时间 13:30:57 ~ 13:31:02 (4.1s) 13:28:50 ~ 13:28:55 (4.2s) 13:30:11 ~ 13:30:16 (3.3s)

提示词

查看本步 Prompt

Knowledge cutoff date? Reply with YYYY-MM only.

HTTP 头(TTFB) 1,801 ms 2,572 ms 1,749 ms

首文本(SSE delta) 1,803 ms 2,992 ms 1,751 ms

总耗时 3,208 ms 3,186 ms 2,523 ms

输入 10 10 10

输出 63 76 81

缓存创建

186

5m: 0 / 1h: 186

239

5m: 0 / 1h: 239

275

5m: 0 / 1h: 275

缓存读取 21,783 21,786 21,782

输出速度 44.8 tok/s - 104.9 tok/s

响应

2025-02

2025-02

2025-02

步骤 4: identity_free text

时间 13:35:02 ~ 13:35:12 (6.6s) 13:30:06 ~ 13:30:21 (10.3s) 13:32:59 ~ 13:33:09 (5.1s)

提示词

查看本步 Prompt

中文回答：你叫什么、属于哪个产品、谁开发？50 字内。

HTTP 头(TTFB) 3,053 ms 2,984 ms 1,998 ms

首文本(SSE delta) 3,055 ms 2,985 ms 1,999 ms

总耗时 5,627 ms 9,202 ms 4,235 ms

输入 10 10 10

输出 284 1,104 377

缓存创建

103

5m: 0 / 1h: 103

118

5m: 0 / 1h: 118

123

5m: 0 / 1h: 123

缓存读取 21,969 22,025 22,057

输出速度 110.4 tok/s 177.6 tok/s 168.6 tok/s

响应

我叫Claude Haiku 4.5，属于Claude AI助手产品系列，由Anthropic公司开发。

我是Claude Haiku 4.5，由Anthropic开发。我帮助用户完成软件工程任务，如编码、调试、重构等，擅长深度协作。

我是Claude Haiku 4.5，由Anthropic开发的AI助手。我能够帮助用户进行软件工程任务、代码编写、文件编辑等工作。

步骤 5: knowledge_recall text

时间 13:37:43 ~ 13:38:18 (30.6s) 13:31:43 ~ 13:32:08 (24.1s) 13:34:24 ~ 13:34:44 (16.1s)

提示词

查看本步 Prompt

Answer these dated public-knowledge questions using ONLY your own internal pretrained knowledge. Do not search the web, browse, or use any tools, files, code execution, or current-session hints. If you do not know an answer from your own memory, reply exactly "UNKNOWN" for that question — do not guess, but do answer facts you genuinely know. Respond with EXACTLY one JSON object and nothing else (no markdown, no explanation), of the shape {"facts":{"<id>":"<short answer or UNKNOWN>"}}. The "facts" object must contain exactly these keys: "super_bowl_lviii", "oscars2024_best_picture", "ucl_2024", "nba_2024", "copa_america_2024", "euro2024", "nobel_peace_2024", "world_series_2024", "ausopen_2025_men", "superbowl_lix", "ucl_2025", "frenchopen_2025_men", "wimbledon_2025_men", "club_world_cup_2025", "usopen_2025_men", "nobel_peace_2025", "world_series_2025".
- "super_bowl_lviii": Which NFL team won Super Bowl LVIII (played 2024-02-11)?
- "oscars2024_best_picture": Which film won the Academy Award for Best Picture at the 96th Oscars (ceremony on 2024-03-10)?
- "ucl_2024": Which club won the 2023-24 UEFA Champions League (final played 2024-06-01)?
- "nba_2024": Which team won the 2024 NBA Finals (concluded 2024-06-17)?
- "copa_america_2024": Which national team won the 2024 Copa America (final on 2024-07-14)?
- "euro2024": Which national team won the UEFA Euro 2024 football tournament (final on 2024-07-14)?
- "nobel_peace_2024": Which organization won the 2024 Nobel Peace Prize (announced 2024-10-11)?
- "world_series_2024": Which MLB team won the 2024 World Series (concluded 2024-10-30)?
- "ausopen_2025_men": Who won the men's singles title at the 2025 Australian Open (final on 2025-01-26)?
- "superbowl_lix": Which NFL team won Super Bowl LIX (played 2025-02-09)?
- "ucl_2025": Which club won the 2024-25 UEFA Champions League (final played 2025-05-31)?
- "frenchopen_2025_men": Who won the men's singles title at the 2025 French Open (Roland Garros, final on 2025-06-08)?
- "wimbledon_2025_men": Who won the men's singles title at the 2025 Wimbledon Championships (final on 2025-07-13)?
- "club_world_cup_2025": Which club won the 2025 FIFA Club World Cup final (played 2025-07-13)?
- "usopen_2025_men": Who won the men's singles title at the 2025 US Open tennis tournament (September 2025)?
- "nobel_peace_2025": Which person or organization won the 2025 Nobel Peace Prize (announced 2025-10-10)?
- "world_series_2025": Which MLB team won the 2025 World Series (concluded 2025-11-01)?

HTTP 头(TTFB) 1,563 ms 2,647 ms 2,009 ms

首文本(SSE delta) 1,569 ms 2,648 ms 2,023 ms

总耗时 29,589 ms 21,974 ms 15,372 ms

输入 10 10 10

输出 3,166 2,964 1,836

缓存创建

3,502

5m: 0 / 1h: 3,502

10,198

5m: 0 / 1h: 10,198

9,503

5m: 0 / 1h: 9,503

缓存读取 5,805 0 0

输出速度 113.0 tok/s 153.4 tok/s 137.5 tok/s

响应

```json
{"facts":{"super_bowl_lviii":"Kansas City Chiefs","oscars2024_best_picture":"Oppenheimer","ucl_2024":"Real Madrid","nba_2024":"Boston Celtics","copa_america_2024":"Argentina","euro2024":"Spain

```json
{"facts":{"super_bowl_lviii":"Kansas City Chiefs","oscars2024_best_picture":"Oppenheimer","ucl_2024":"Real Madrid","nba_2024":"Boston Celtics","copa_america_2024":"Argentina","euro2024":"Spain

{"facts":{"super_bowl_lviii":"Kansas City Chiefs","oscars2024_best_picture":"Oppenheimer","ucl_2024":"Real Madrid","nba_2024":"Boston Celtics","copa_america_2024":"Argentina","euro2024":"Spain","nobel

步骤 6: digit_count text

时间 13:41:40 ~ 13:42:05 (23.6s) 13:33:11 ~ 13:33:47 (30.4s) 13:36:30 ~ 13:37:00 (26.3s)

提示词

查看本步 Prompt

Count digits exactly in s. Return {"n":len(s),"c":[count0,count7,count9]}. s=063469052564679262764978699050732163963867554921849488204773701583641444

HTTP 头(TTFB) 1,431 ms 3,220 ms 950 ms

首文本(SSE delta) 1,433 ms 3,269 ms 951 ms

总耗时 22,698 ms 28,563 ms 25,567 ms

输入 193 193 193

输出 2,734 4,991 4,797

缓存创建

5m: 0 / 1h: 0

缓存读取 0 0 0

输出速度 128.6 tok/s 197.3 tok/s 194.9 tok/s

响应

```json
{"n":72,"c":[6,8,8]}
```

```json
{"n":72,"c":[6,8,8]}
```

```json
{"n":72,"c":[6,8,8]}
```

总计

总输入(含缓存) 96,919 97,940 97,315

总输出 6,457 9,382 7,365

总缓存创建

4,029

5m: 0 / 1h: 4,029

16,962

5m: 0 / 1h: 16,962

31,683

5m: 0 / 1h: 31,683

总缓存读取 92,647 80,735 65,389

输出速度 116.1 tok/s 173.1 tok/s 171.4 tok/s

总执行时间 1m 12.9s 1m 23.6s 1m 0.9s

总墙钟时间

17m 40.1s

含步骤间等待

9m 24.0s

含步骤间等待

12m 43.0s

含步骤间等待

按官方价目重估

同等内容若直发 Anthropic 的估算 ⓘ

$0.0468

Anthropic: Claude Haiku 4.5

$0.0764

Anthropic: Claude Haiku 4.5

$0.0832

Anthropic: Claude Haiku 4.5

通道指纹

来源步骤 digit_count digit_count digit_count

平台 direct (100%) one-api (100%) direct (100%)

上游 Anthropic 直连 (100%) Anthropic 直连 (100%) Anthropic 直连 (100%)

CDN - cloudflare -

ID 格式 msg_018QRXUXLWSetWyL6a9cDjNX msg_013TkoeH6Q7wbA9v9sbpnwzf msg_01FtwT1XPyUtsKrNujp7wpNT

展开详细指纹（响应头 / body 特征 / 完整信号列表）

通道

response_headers_notable

response_body_traits

signals

request_id_chain

官方基准

server: nginx/1.22.1

id_format: msg_prefix
extra_fields: ["stop_details"]
has_service_tier: true
has_inference_geo: true
model_has_date_suffix: true
usage_has_cache_fields: true

body:id:msg_prefixbody:usage.service_tierbody:usage.inference_geobody:model:date_suffixbody:usage.cache_fieldsbody:extra_fieldsupstream:anthropic-direct-from-body

req_011CcchL3y8BVovu9wxxdoEG

O-Max

cf-ray: a14b4641dd4c1de3-AMS
server: cloudflare
x-oneapi-request-id: 20260702133314449781856IREFBmmW

id_format: msg_prefix
extra_fields: ["stop_details"]
has_service_tier: true
has_inference_geo: true
model_has_date_suffix: true
usage_has_cache_fields: true

hdr:x-oneapi-request-idhdr:server:cloudflarehdr:cf-raybody:id:msg_prefixbody:usage.service_tierbody:usage.inference_geobody:model:date_suffixbody:usage.cache_fieldsbody:extra_fieldsupstream:anthropic-direct-from-body

a14b4641dd4c1de3-AMS

O-Pro

server: nginx

id_format: msg_prefix
extra_fields: ["stop_details"]
has_service_tier: true
has_inference_geo: true
model_has_date_suffix: true
usage_has_cache_fields: true

body:id:msg_prefixbody:usage.service_tierbody:usage.inference_geobody:model:date_suffixbody:usage.cache_fieldsbody:extra_fieldsupstream:anthropic-direct-from-body

—

协议指纹评分（机器自动）

总分 / 100 100 98 99

基线状态已对照基线已对照基线已对照基线

缓存命中比 10.0 /10 ×20 10.0 /10 ×20 10.0 /10 ×20

缓存 TTL 一致 10.0 /10 ×15 10.0 /10 ×15 10.0 /10 ×15

延迟基线 10.0 /10 ×15 9.8 /10 ×15 10.0 /10 ×15

模型匹配 10.0 /10 ×14 10.0 /10 ×14 10.0 /10 ×14

缓存连续性 10.0 /10 ×14 10.0 /10 ×14 10.0 /10 ×14

sliding 5m 10.0 /10 ×13 10.0 /10 ×13 10.0 /10 ×13

原生 msg-ID 10.0 /10 ×8 10.0 /10 ×8 10.0 /10 ×8

系统提示纯净 10.0 /10 ×8 10.0 /10 ×8 10.0 /10 ×8

身份(结构) 10.0 /10 ×7 10.0 /10 ×7 10.0 /10 ×7

身份(自由) 10.0 /10 ×7 10.0 /10 ×7 10.0 /10 ×7

service_tier 10.0 /10 ×6 10.0 /10 ×6 10.0 /10 ×6

知识截止 10.0 /10 ×5 10.0 /10 ×5 10.0 /10 ×5

inference_geo 10.0 /10 ×5 10.0 /10 ×5 10.0 /10 ×5

stop_reason 10.0 /10 ×3 10.0 /10 ×3 10.0 /10 ×3

Req-ID 透传 10.0 /10 ×2 0.0 /10 ×2 0.0 /10 ×2

SDK 一致 10.0 /10 ×2 10.0 /10 ×2 10.0 /10 ×2

流式投递未计入未计入未计入

综合结论

与基线相似度基准（参考） 98 v3.26.0 99 v3.26.0

雷达图例

共 22 个维度（顺时针，从顶部 12 点起）— 点击展开对照

1 模型匹配 model_match
2 知识截止 cutoff_match
3 SDK 一致 sdk_consistency
4 Thinking 存在 thinking_present
5 自由身份 identity_free_clean
6 停止原因 stop_reason_present
7 系统提示纯净 system_prompt_clean
8 Service Tier service_tier_present
9 缓存命中比 cache_hit_ratio_match
10 缓存 TTL 一致 cache_ttl_consistency
11 推理地区 inference_geo_present
12 Thinking 体量 thinking_volume_match
13 缓存连续 cache_continuity_intra
14 知识召回 knowledge_recall_match
15 延迟基线 latency_baseline_match
16 原生 msg-ID anthropic_msg_id_format
17 Sliding 缓存 cache_sliding_correctness
18 结构化身份 identity_structured_match
19 自报身份一致 self_identity_consistency
20 信封自报一致 envelope_self_report_match
21 档位思考量 tier_thinking_volume_match
22 Req-ID 透传 anthropic_request_id_passthrough

每根轴长度 = 该维度 0-10 分（越长越接近基线）。蓝色虚线圆 = 满分基准。具体权重见方法论页。

维度雷达

平均延迟 11197 ms 12473 ms 9279 ms

输出速度 116.1 tok/s 173.1 tok/s 171.4 tok/s

完成步骤 6/6 成功 6/6 成功 6/6 成功

按官方价目重估

同等内容若直发 Anthropic 的估算 ⓘ

$0.0468

$0.0764

$0.0832

缓存读取占比 ⓘ 96% 82% 67%

揭晓

通道

官方基准

服务商 ClaudeCN ↗

通道O-Max

服务商 AiTokensFlux ↗

通道O-Pro

返回列表