Free model verification · Redacted public report

claude-sonnet-4-6 standard verification model-consistency report

Report ID MZZR20260629093652FDD55DGenerated at 2026/06/29 17:36

Doubtful

claude-sonnet-4-6 本次模型一致性标准验真结果：Doubtful（68/100）。主要扣分项：模型代码锚定不稳定；指令跟随不稳定；模型族识别不稳定。

68Consistency score

主要扣分项：模型代码锚定不稳定；指令跟随不稳定；模型族识别不稳定。

EvaluatorMozhenzhen

Tested modelClaude Sonnet 4.6

Providersiksvip

Samples22 probes

Duration2 min 53 sec

Verification versionV10

Identity verdictNot verified

Verification basisMozhenzhen versioned baseline

Token usage

Measured API usage is shown first; official baselines or test estimates are used only when measured values are unavailable.

Measured token usage59.7KFrom the response usage fields; this is the report's primary usage measure.Test-suite estimate74.0KPrecisely estimated from this test suite when no official baseline is available.Variance assessment19% belowMeasured token usage is below the estimate, possibly due to tokenizer, compression, or usage-reporting differences.Usage conclusionBelow estimateThis reflects usage observability, not the final billed amount.

Token cache usage test

Observes token usage, cache fields, and reuse across rounds; 0% means no cached-token field was observed.

Tested 5 times

Observed cache share60.4%

Low cache share

Cached tokens were observed, but the share is low. Review API parameters, context reuse, and billing methodology.

Observed tokens50.5KCached tokens30.5KNon-cached tokens20.0KAverage per run10.1K

Token usage recordReturnedCache fieldObservedResultNormal

Cache share by testRepresents the API-reported cache share, not the final billing discount.

Long-context test

Shows stable responses at different context lengths and the model center reference window.

2 samples

Verified up to 8K

16K 需要关注，原因：上下文保持较弱。

Reference windowReference window 1MModel-center context window; available lengths may vary by provider.Tested context lengths8K / 16KTests begin at the longest generated context and step down to verify stable, correct responses.

Current guidance

16K 档暂不作为当前线路的稳定上下文参考。

Model authenticity score

Review API availability, model identity, response completeness, and other checks separately.

Scoring breakdown

Doubtful68/100

主要扣分项：模型代码锚定不稳定；指令跟随不稳定；模型族识别不稳定。

Callability and model visibility15 / 15

No deductions

Protocol and response structure15 / 15

No deductions

Identity and model match0 / 20

品牌边界表达不清（品牌题：Claude 思考边界） -4模型代码锚定不稳定（模型级题：模型代码锚定） -10模型族识别不稳定（模型族题：系列和档位识别） -6

能力题表现9 / 20

指令跟随不稳定（标准题：指令跟随） -8模型代码锚定不稳定（模型级题：模型代码锚定） -3

Usage and billing consistency15 / 15

No deductions

稳定性与复测表现14 / 15

多个关键证据项需要复核 -1

Test details

Expand each check to view the redacted evidence notes.

13 checks

Model list visibility

claude-sonnet-4-6 is visible in the API model list.

Passed

Test evidence

Check: Model list visibility
Status: Passed
Samples: 6
API path: /v1/models
Visible models: 6

Basic call probe

The basic API call succeeded and returned a recognizable response.

Passed

Test evidence

Check: Basic call probe
Status: Passed
Samples: 1
API path: /v1/chat/completions

Quick test: model self-identification

直接询问模型身份，作为低权重红旗信号；自报只能提示明显不一致，不能单独证明官方身份。全部样本Passed。

Passed

Test evidence

Check: Quick test: model self-identification
Status: Passed
Samples: 1
Test type: Text test
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 155

Standard test: instruction following

要求模型按指定标记回复，判断基础指令跟随和响应可控性。存在样本未完全Passed。

Needs review

Test evidence

Check: Standard test: instruction following
Status: Needs review
Samples: 1
Test type: Text test
Needs review原因：指令跟随不稳定
Results: Passed 0 / Needs review 1 / Failed 0

Standard test: JSON structure

要求模型输出固定 JSON，判断结构化响应是否稳定。全部样本Passed。

Passed

Test evidence

Check: Standard test: JSON structure
Status: Passed
Samples: 1
Test type: Text test
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 487

Standard test: brand boundary

要求模型不要伪造官方身份，判断品牌表述是否克制。全部样本Passed。

Passed

Test evidence

Check: Standard test: brand boundary
Status: Passed
Samples: 1
Test type: Text test
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 182

品牌题：Claude 思考边界

观察模型是否会克制说明隐藏思考和签名特征不能被随意伪造。存在样本未完全Passed。

Needs review

Test evidence

Check: 品牌题：Claude 思考边界
Status: Needs review
Samples: 1
Test type: Text test
Needs review原因：品牌边界表达不清
Results: Passed 0 / Needs review 1 / Failed 0

Model-specific test: model-code anchoring

要求模型把本次被检测模型代码、品牌归类和固定标记写入 JSON，便于应用层精确校验。存在样本未完全Passed。

Needs review

Test evidence

Check: Model-specific test: model-code anchoring
Status: Needs review
Samples: 1
Test type: Text test
Needs review原因：模型代码锚定不稳定
Results: Passed 0 / Needs review 1 / Failed 0

Model-family test: series and tier identification

按模型代码识别系列或档位，覆盖 Opus/Sonnet/Haiku、GPT/o/Codex、Gemini Pro/Flash、GLM、Kimi、MiniMax 等模型族。存在样本未完全Passed。

Needs review

Test evidence

Check: Model-family test: series and tier identification
Status: Needs review
Samples: 1
Test type: Text test
Needs review原因：模型族识别不稳定
Results: Passed 0 / Needs review 1 / Failed 0

Standard test: usage observability

要求短输出并检查响应 usage 是否可被应用层读取，支撑预计消耗与实际消耗对照。全部样本Passed。

Passed

Test evidence

Check: Standard test: usage observability
Status: Passed
Samples: 1
Test type: Text test
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 500

Specialty test: token cache usage over 5 rounds

同一长前缀连续请求 5 次，应用层观察 usage、cache usage、token 消耗和费用风险。全部样本Passed。

Passed

Test evidence

Check: Specialty test: token cache usage over 5 rounds
Status: Passed
Samples: 5
Test type: Text test
Specialty: Token cache usage test (may affect real billing)
Results: Passed 5 / Needs review 0 / Failed 0
Observed tokens: 50.5K
Cached tokens: 30.5K

专项：8K 上下文能力实测

8K context sample returned reliably and is included as evidence for this run.

Passed

Test evidence

Check: 专项：8K 上下文能力实测
Status: Passed
Samples: 1
Test type: Text test
Specialty: 上下文能力实测
Context tier: 8K
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 7.8K

专项：16K 上下文能力实测

16K 长度样本未形成稳定结果，本次按边界样本记录。

Needs review

Test evidence

Check: 专项：16K 上下文能力实测
Status: Needs review
Samples: 1
Test type: Text test
Specialty: 上下文能力实测
Context tier: 16K
Needs review原因：上下文保持较弱
Results: Passed 0 / Needs review 1 / Failed 0

Report comments

No public reviews

Be the first to leave a review.

Post an anonymous comment

Rating

5 points

This reportReviews

Upload images/videoClick or drag to uploadUp to 9 images and 1 video · Sign-in required for attachments

AI service research and model verification report

Scan to view reportOpen full report

Free model verification · Supports all text models

claude-sonnet-4-6 standard verification model-consistency report

Report ID:MZZR20260629093652FDD55DGenerated at：2026/06/29 17:36

Provider details AI model details

Model consistency score

Doubtful

claude-sonnet-4-6 本次模型一致性标准验真结果：Doubtful（68/100）。主要扣分项：模型代码锚定不稳定；指令跟随不稳定；模型族识别不稳定。

EvaluatorMozhenzhen

Tested modelClaude Sonnet 4.6

Providersiksvip

Samples22

Duration3 min

Test versionV10

Identity verdictNot verified

Verification basisMozhenzhen versioned baseline

Token usage

Measured API usage is shown first; official baselines or test estimates are used only when measured values are unavailable.

Token cache usage test

Observes token usage, cache fields, and reuse across rounds; 0% means no cached-token field was observed.

Tested 5 times

Observed cache share60.4%

Low cache share

Cached tokens were observed, but the share is low. Review API parameters, context reuse, and billing methodology.

Observed tokens50.5KCached tokens30.5KNon-cached tokens20.0KAverage per run10.1K

Token usage recordReturnedCache fieldObservedResultNormal

Cache share by testRepresents the API-reported cache share, not the final billing discount.

Long-context test

Shows stable responses at different context lengths and the model center reference window.

2 samples

Verified up to 8K

16K 需要关注，原因：上下文保持较弱。

Current guidance

16K 档暂不作为当前线路的稳定上下文参考。

Model authenticity score

Review API availability, model identity, response completeness, and other checks separately.

Scoring breakdown

Doubtful68/100

主要扣分项：模型代码锚定不稳定；指令跟随不稳定；模型族识别不稳定。

Callability and model visibility15 / 15

No deductions

Protocol and response structure15 / 15

No deductions

Identity and model match0 / 20

能力题表现9 / 20

指令跟随不稳定（标准题：指令跟随） -8模型代码锚定不稳定（模型级题：模型代码锚定） -3

Usage and billing consistency15 / 15

No deductions

稳定性与复测表现14 / 15

多个关键证据项需要复核 -1

Test details

Expand each check to view the redacted evidence notes.

13 checks

Model list visibility

claude-sonnet-4-6 is visible in the API model list.

Passed

Test evidence

Check: Model list visibility
Status: Passed
Samples: 6
API path: /v1/models
Visible models: 6

Basic call probe

The basic API call succeeded and returned a recognizable response.

Passed

Test evidence

Check: Basic call probe
Status: Passed
Samples: 1
API path: /v1/chat/completions

Quick test: model self-identification

直接询问模型身份，作为低权重红旗信号；自报只能提示明显不一致，不能单独证明官方身份。全部样本Passed。

Passed

Test evidence

Check: Quick test: model self-identification
Status: Passed
Samples: 1
Test type: Text test
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 155

Standard test: instruction following

要求模型按指定标记回复，判断基础指令跟随和响应可控性。存在样本未完全Passed。

Needs review

Test evidence

Check: Standard test: instruction following
Status: Needs review
Samples: 1
Test type: Text test
Needs review原因：指令跟随不稳定
Results: Passed 0 / Needs review 1 / Failed 0

Standard test: JSON structure

要求模型输出固定 JSON，判断结构化响应是否稳定。全部样本Passed。

Passed

Test evidence

Check: Standard test: JSON structure
Status: Passed
Samples: 1
Test type: Text test
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 487

Standard test: brand boundary

要求模型不要伪造官方身份，判断品牌表述是否克制。全部样本Passed。

Passed

Test evidence

Check: Standard test: brand boundary
Status: Passed
Samples: 1
Test type: Text test
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 182

品牌题：Claude 思考边界

观察模型是否会克制说明隐藏思考和签名特征不能被随意伪造。存在样本未完全Passed。

Needs review

Test evidence

Check: 品牌题：Claude 思考边界
Status: Needs review
Samples: 1
Test type: Text test
Needs review原因：品牌边界表达不清
Results: Passed 0 / Needs review 1 / Failed 0

Model-specific test: model-code anchoring

要求模型把本次被检测模型代码、品牌归类和固定标记写入 JSON，便于应用层精确校验。存在样本未完全Passed。

Needs review

Test evidence

Check: Model-specific test: model-code anchoring
Status: Needs review
Samples: 1
Test type: Text test
Needs review原因：模型代码锚定不稳定
Results: Passed 0 / Needs review 1 / Failed 0

Model-family test: series and tier identification

按模型代码识别系列或档位，覆盖 Opus/Sonnet/Haiku、GPT/o/Codex、Gemini Pro/Flash、GLM、Kimi、MiniMax 等模型族。存在样本未完全Passed。

Needs review

Test evidence

Check: Model-family test: series and tier identification
Status: Needs review
Samples: 1
Test type: Text test
Needs review原因：模型族识别不稳定
Results: Passed 0 / Needs review 1 / Failed 0

Standard test: usage observability

要求短输出并检查响应 usage 是否可被应用层读取，支撑预计消耗与实际消耗对照。全部样本Passed。

Passed

Test evidence

Check: Standard test: usage observability
Status: Passed
Samples: 1
Test type: Text test
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 500

Specialty test: token cache usage over 5 rounds

同一长前缀连续请求 5 次，应用层观察 usage、cache usage、token 消耗和费用风险。全部样本Passed。

Passed

Test evidence

Check: Specialty test: token cache usage over 5 rounds
Status: Passed
Samples: 5
Test type: Text test
Specialty: Token cache usage test (may affect real billing)
Results: Passed 5 / Needs review 0 / Failed 0
Observed tokens: 50.5K
Cached tokens: 30.5K

专项：8K 上下文能力实测

8K context sample returned reliably and is included as evidence for this run.

Passed

Test evidence

Check: 专项：8K 上下文能力实测
Status: Passed
Samples: 1
Test type: Text test
Specialty: 上下文能力实测
Context tier: 8K
Results: Passed 1 / Needs review 0 / Failed 0
Observed tokens: 7.8K

专项：16K 上下文能力实测

16K 长度样本未形成稳定结果，本次按边界样本记录。

Needs review

Test evidence

Check: 专项：16K 上下文能力实测
Status: Needs review
Samples: 1
Test type: Text test
Specialty: 上下文能力实测
Context tier: 16K
Needs review原因：上下文保持较弱
Results: Passed 0 / Needs review 1 / Failed 0

Latest user reviews

No public reviews

Be the first to leave a review.

Comments and feedback on this report

Rating

5 points

This reportReviews

Upload images/videoClick or drag to uploadUp to 9 images and 1 video · Sign-in required for attachments