Torna alle skill
folkbench
di Folkbench
Aider—Non ancora valutato
La tua valutazione
Sintesi
AI API 中转站推荐与评测:中转站排行榜、中转站检测、Claude / GPT / Gemini 中转站对比 · AI API relay rankings · Folkbench Switch for Claude Code, Codex & Gemini CLI
Descrizione
<div align="center">
<img src="docs/images/readme-zh/hero.webp" width="960" alt="Folkbench:在真实场景中衡量 AI 的价值。公开评测、降智监控与一键切换。">
<br>
[](https://github.com/Folkbench/folkbench/releases)
[](https://github.com/Folkbench/folkbench/stargazers)
[](https://folkbench.com/?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-home-1)
**[查看榜单](https://folkbench.com/rankings?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-rankings-2)** · **[降智监控](https://folkbench.com/arena?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-arena-3)** · **[下载 Switch](https://github.com/Folkbench/folkbench/releases)**
[English](README.en.md) · [日本語](README.ja.md) · [Español](README.es.md)
</div>
<br>
## 中转站推荐
以下中转站支持 Folkbench 的持续运营。
<table>
<tr><td width="180"><a href="https://folkbench.com/go/modelflare?source=readme"><picture><source media="(prefers-color-scheme: dark)" srcset="docs/images/sponsors/modelflare-dark.png"><img src="docs/images/sponsors/modelflare-light.png" alt="Modelflare" width="160"></picture></a></td><td><a href="https://folkbench.com/go/modelflare?source=readme">Modelflare</a><br>Modelflare站点是由美国 Havenbyte LLC 运营,支持 GPT、Claude、Gemini、Grok 等主流模型和常用兼容协议,同时也支持 40 多种图片及视频模型。API Key 可以配置有序回退,一把 API Key 可以调用多个模型,Key 还可以设置备用分组,并且从请求日志里能看到状态、延迟、Token 和实际费用。<br><br>缓存补偿是这个站的另一个亮点:部分 OpenAI 价格优先、稳定和高级分组分别提供每日 65%、75%、85% 的缓存命中率保障,对于符合条件的api请求如果未达到缓存命中率,平台会给予对应补充。<br><br>站点目前也有首次充值解锁特惠分组的活动,单笔充值 US$20 后可解锁 0.015 倍率 GPT 特惠分组,解锁后 GPT-5.6-Sol 的价格大约是 ¥0.51(输入)/ ¥3.01(输出)/每百万 Token,并且没有使用上限限制。单笔充值 US$50 后可解锁 0.03 倍率 Claude 特惠分组。</td></tr>
<tr><td width="180"><a href="https://folkbench.com/go/ikuncode?source=readme"><picture><source media="(prefers-color-scheme: dark)" srcset="docs/images/sponsors/ikuncode-dark.png"><img src="docs/images/sponsors/ikuncode-light.png" alt="IKunCode" width="160"></picture></a></td><td><a href="https://folkbench.com/go/ikuncode?source=readme">IKunCode</a><br>IKunCode对于各个编程工具接入的支持比较好,Claude Code、Codex 和 Gemini CLI 都有单独支持,售后以及咨询渠道渠道做得比较全,官网公开了客服邮箱、QQ 群和 Telegram,人工客服服务贴心,很适合新手上路。</td></tr>
<tr><td width="180"><a href="https://folkbench.com/go/fluxlane?source=readme"><picture><source media="(prefers-color-scheme: dark)" srcset="docs/images/sponsors/fluxlane-dark.png"><img src="docs/images/sponsors/fluxlane-light.png" alt="智流 FluxLane" width="160"></picture></a></td><td><a href="https://folkbench.com/go/fluxlane?source=readme">智流 FluxLane</a><br>智流 FluxLane 是一个面向开发者与团队的 AI API 网关:提供满血且稳定的全球模型,支持 GPT、Claude、Gemini、Deepseek、GLM 等主流模型和全球主流模型厂商协议,平台提供策略路由与多线路容灾,方便国内团队长期接入与成本核算。面向国内企业/高校采购,可签订合同、对公转账,并提供发票、采购单等材料(具体以商务确认为准),公开模型能力检测报告,倍率低至0.1倍率,新人注册可领 2 美元体验额度。</td></tr>
<tr><td width="180"><a href="https://folkbench.com/go/merchant-955eb49b-34a?source=readme"><picture><source media="(prefers-color-scheme: dark)" srcset="docs/images/sponsors/sidrune-dark.png"><img src="docs/images/sponsors/sidrune-light.png" alt="Sidrune AI" width="160"></picture></a></td><td><a href="https://folkbench.com/go/merchant-955eb49b-34a?source=readme">Sidrune AI</a><br>省心用Claude:Claude Max低至1.5折,另有14+Claude API线路,按量计费,支持开票及对公转账。</td></tr>
</table>
## AI API 中转站推荐与选站指南
> 排序遵循 Folkbench 站点规则,包含合作等级和运营加权;部分数据为展示样例。
### 中转站怎么选
1. **模型是否降级**:核对中转站的模型与分组,用相同任务比较回答、长文本处理和工具调用表现;能返回内容不代表模型符合预期。
2. **可用率**:查看中转站的统计时段与请求数量,同时留意空回复、超时和流式中断,小样本的高可用率不能代表长期表现。
3. **价格倍率**:比较中转站报价时对齐模型、币种与计价单位,分别看输入、输出和缓存费用,再核对实际扣费。
4. **退款政策**:先看中转站是否写清退款条件、手续费、处理时间与联系渠道;条款不明确时,先问清再充值。
5. **先小额试用**:用中转站的最低充值档或试用额度跑自己的常用任务,确认客户端兼容、响应与账单符合需求后再决定用量。
### 中转站充值风险
- **不要大额充值**:按近期用量控制中转站余额,避免为了折扣长期囤款。
- **留意跑路信号**:中转站突然高额返利、客服长期失联、退款受阻或频繁换域名时,先暂停追加充值并保存交易记录。
- **不要外贴 Key**:向中转站反馈故障或在 GitHub 提交问题时,遮盖 API Key、令牌和含凭据的截图。
- **留意数据去向**:提交内容会经过中转站;使用前确认数据保留和处理条款,避免发送不必要的敏感资料。
### 怎么检测一个中转站有没有降智
1. 记录中转站地址、模型、分组与测试时间,使用日常客户端检查流式响应和工具调用。
2. 选几道有明确预期的推理、代码和长文本任务,在中转站与官方接口上使用相同参数,多次比较结果。
3. 到 [Folkbench 中转站降智监控](https://folkbench.com/arena?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=relay-guide-arena) 查看已有探针结果,并结合[评测方法](https://folkbench.com/methodology?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=relay-guide-methodology)理解检测条件和局限。中转站的一次异常回答不足以证明模型被替换。
### Claude Code / Codex / Gemini CLI 中转站怎么选
Folkbench 帮你挑线路、配好客户端,不是自建代理教程。
- **Claude Code 中转站:** 先看 [Claude 中转站测评](https://folkbench.com/claude-zhongzhuanzhan-ceping?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=tool-claude-ceping),对比可用率、首字延迟和会话成本。
- **Codex 中转站:** 先看 [GPT / Codex 中转站测评](https://folkbench.com/gpt-codex-zhongzhuanzhan-ceping?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=tool-codex-ceping)。
- **一键切换配置:** 选好线路后,用 [Folkbench Switch](https://folkbench.com/switch?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=tool-switch) 在 Claude Code、Codex 和 Gemini CLI 之间切换中转站配置。
- **中转站横向对比:** 想把几家中转站放在一起看,用 [Folkbench 中转站对比](https://folkbench.com/stations/compare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=tool-compare) 并排比较。排序按站点规则展示,包含合作档位和运营加权;部分数据为展示样本。
### 中转站排行榜
查看 [Folkbench 中转站排行榜](https://folkbench.com/rankings?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=relay-guide-rankings),或阅读完整的[中转站推荐清单(GitHub)](https://github.com/magnus-minor/relay-api-recommend),结合自己的模型需求与小额试用结果选择。
想先自己验一下手上的中转站?用 [Folkbench 实时检测](https://folkbench.com/evaluations/api-quality?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=relay-guide-check),填入你自己的 Key 检测一个接口:报告仅自己可见,密钥不保留。
## 今日公开快照
<!-- FOLKBENCH-BOARD:START -->
> 归档日期:2026-11-21 · 发布:2026-11-21 20:30:55(北京时间)· 快照:`public-162129ede5b493c61812986fef1a8b5c30cd3bc0c8e5efd0fe3b4d7b942b1eb3`
> 排序沿用主站规则,包含合作档位与运营加分;部分数据为展示样本。本页提供公开快照,不是原始运行记录;部分数据为展示样本。
| 模型 | # | 中转站 | 综合 | 可用率 | 稳定性 | 输入 / 输出价 | 样本 |
| :-- | :-: | :-- | --: | --: | --: | :-- | --: |
| Claude Fable 5.1 | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-fable-5-1-r1) | 100.0% | 97.6% | 95.0% | ¥1.99 / 1M / ¥9.96 / 1M | 168 |
| Claude Fable 5.1 | 2 | [Apiporter](https://folkbench.com/services/apiporter?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-fable-5-1-r2) | 95.2% | 100.0% | 85.2% | ¥20.00 / 1M / ¥100.00 / 1M | 22 |
| Claude Fable 5.1 | 3 | [Cloudbridge](https://folkbench.com/services/cloudbridge?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-fable-5-1-r3) | 87.2% | 97.9% | 79.7% | ¥31.20 / 1M / ¥156.00 / 1M | 48 |
| | | [完整 Claude Fable 5.1 榜单 →](https://folkbench.com/rankings/claude-fable-5-1?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-fable-5-1-more) | | | | | |
| Claude Haiku 5.5 | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-haiku-5-5-r1) | 100.0% | 97.6% | 95.3% | ¥0.17 / 1M / ¥0.83 / 1M | 168 |
| claude-haiku-5-5 | 2 | [ModelPod](https://folkbench.com/services/merchant-27af1806-16e?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-haiku-5-5-r2) | 94.0% | 100.0% | 86.7% | ¥0.25 / 1M / ¥1.26 / 1M | 38 |
| claude-haiku-5-5 | 3 | [Sidrune AI](https://folkbench.com/services/merchant-955eb49b-34a?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-haiku-5-5-r3) | 93.5% | 100.0% | 91.3% | ¥0.10 / 1M / ¥0.50 / 1M | 48 |
| | | [完整 Claude Haiku 5.5 榜单 →](https://folkbench.com/rankings/claude-haiku-5-5?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-haiku-5-5-more) | | | | | |
| Claude Opus 5 | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-opus-5-r1) | 100.0% | 98.2% | 95.3% | ¥1.00 / 1M / ¥4.98 / 1M | 165 |
| | | [完整 Claude Opus 5 榜单 →](https://folkbench.com/rankings/claude-opus-5?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-opus-5-more) | | | | | |
| Claude Opus 5.5 | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-opus-5-5-r1) | 100.0% | 98.1% | 94.6% | ¥0.80 / 1M / ¥3.98 / 1M | 104 |
| Claude Opus 5.5 | 2 | [IKunCode](https://folkbench.com/services/ikuncode?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-opus-5-5-r2) | 95.5% | 100.0% | 88.8% | ¥4.00 / 1M / ¥20.00 / 1M | 48 |
| Claude Opus 5.5 | 3 | [智流 FluxLane](https://folkbench.com/services/fluxlane?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-opus-5-5-r3) | 93.5% | 97.9% | 84.0% | ¥4.80 / 1M / ¥24.00 / 1M | 48 |
| | | [完整 Claude Opus 5.5 榜单 →](https://folkbench.com/rankings/claude-opus-5-5?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-claude-opus-5-5-more) | | | | | |
| DeepSeek V4.1 Flash | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-deepseek-v4-1-flash-r1) | 100.0% | 97.7% | 95.1% | ¥1.20 / 1M / ¥4.80 / 1M | 170 |
| DeepSeek V4.1 Flash | 2 | [IKunCode](https://folkbench.com/services/ikuncode?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-deepseek-v4-1-flash-r2) | 95.5% | 100.0% | 85.5% | — / — | 48 |
| deepseek-v4.1-flash | 3 | [破晓](https://folkbench.com/services/merchant-4cf2b788-920?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-deepseek-v4-1-flash-r3) | 93.6% | 100.0% | 91.3% | ¥0.60 / 1M / ¥2.40 / 1M | 48 |
| | | [完整 DeepSeek V4.1 Flash 榜单 →](https://folkbench.com/rankings/deepseek-v4-1-flash?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-deepseek-v4-1-flash-more) | | | | | |
| Gemini 3.8 Flash | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gemini-3-8-flash-r1) | 100.0% | 97.0% | 94.5% | ¥1.49 / 1M / ¥7.47 / 1M | 168 |
| | | [完整 Gemini 3.8 Flash 榜单 →](https://folkbench.com/rankings/gemini-3-8-flash?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gemini-3-8-flash-more) | | | | | |
| GLM 5.3 | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-glm-5-3-r1) | 100.0% | 97.4% | 94.8% | ¥4.00 / 1M / ¥14.00 / 1M | 192 |
| | | [完整 GLM 5.3 榜单 →](https://folkbench.com/rankings/glm-5-3?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-glm-5-3-more) | | | | | |
| GPT 5.6 Terra | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-5-6-terra-r1) | 100.0% | 97.2% | 94.2% | ¥0.20 / 1M / ¥1.20 / 1M | 145 |
| | | [完整 GPT 5.6 Terra 榜单 →](https://folkbench.com/rankings/gpt-5-6-terra?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-5-6-terra-more) | | | | | |
| GPT 6.1 Sol | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-6-1-sol-r1) | 100.0% | 97.3% | 94.4% | ¥0.20 / 1M / ¥1.00 / 1M | 183 |
| gpt-6.1-sol | 2 | [Sidrune AI](https://folkbench.com/services/merchant-955eb49b-34a?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-6-1-sol-r2) | 93.5% | 100.0% | 84.8% | ¥0.20 / 1M / ¥1.00 / 1M | 48 |
| gpt-6.1-sol | 3 | [Zashmodel](https://folkbench.com/services/merchant-5cf7678f-611?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-6-1-sol-r3) | 87.9% | 100.0% | 75.0% | — / — | 48 |
| | | [完整 GPT 6.1 Sol 榜单 →](https://folkbench.com/rankings/gpt-6-1-sol?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-6-1-sol-more) | | | | | |
| GPT 6 Astra | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-6-astra-r1) | 100.0% | 97.1% | 93.5% | ¥1.00 / 1M / ¥4.98 / 1M | 104 |
| GPT 6 Astra | 2 | [IKunCode](https://folkbench.com/services/ikuncode?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-6-astra-r2) | 95.5% | 100.0% | 92.0% | ¥7.00 / 1M / ¥35.00 / 1M | 48 |
| gpt-6-astra | 3 | [破晓](https://folkbench.com/services/merchant-4cf2b788-920?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-6-astra-r3) | 93.6% | 100.0% | 73.5% | ¥0.50 / 1M / ¥2.50 / 1M | 48 |
| | | [完整 GPT 6 Astra 榜单 →](https://folkbench.com/rankings/gpt-6-astra?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-gpt-6-astra-more) | | | | | |
| Grok 4.6 | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-grok-4-6-r1) | 100.0% | 97.3% | 94.8% | ¥0.13 / 1M / ¥0.40 / 1M | 187 |
| Grok 4.6 | 2 | [IKunCode](https://folkbench.com/services/ikuncode?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-grok-4-6-r2) | 95.5% | 100.0% | 90.2% | ¥0.60 / 1M / ¥1.80 / 1M | 48 |
| | | [完整 Grok 4.6 榜单 →](https://folkbench.com/rankings/grok-4-6?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-grok-4-6-more) | | | | | |
| Grok 4.7 | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-grok-4-7-r1) | 100.0% | 99.0% | 96.5% | ¥1.46 / 1M / ¥4.38 / 1M | 193 |
| grok-4.7 | 2 | [IKunCode](https://folkbench.com/services/ikuncode?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-grok-4-7-r2) | 95.5% | 100.0% | 77.6% | ¥0.60 / 1M / ¥1.80 / 1M | 48 |
| Grok 4.7 | 3 | [Cloudbridge](https://folkbench.com/services/cloudbridge?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-grok-4-7-r3) | 87.2% | 95.8% | 71.5% | ¥1.56 / 1M / ¥4.68 / 1M | 48 |
| | | [完整 Grok 4.7 榜单 →](https://folkbench.com/rankings/grok-4-7?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-grok-4-7-more) | | | | | |
| Qwen 3.8 Max | 1 | [Modelflare](https://folkbench.com/services/modelflare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-qwen3-8-max-r1) | 100.0% | 98.8% | 96.0% | ¥9.60 / 1M / ¥28.80 / 1M | 164 |
| | | [完整 Qwen 3.8 Max 榜单 →](https://folkbench.com/rankings/qwen3-8-max?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-qwen3-8-max-more) | | | | | |
<details open><summary>各模型的来源窗口与版本</summary>
| 模型 | 来源窗口 | 方法版本 | 来源快照 | 发布时间(北京时间) |
| :-- | :-- | :-- | :-- | :-- |
| claude-fable-5-1 | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-8e6db766-1795262739352 | 2026-11-21 20:05:39 |
| claude-haiku-5-5 | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-6b833159-1795023199199 | 2026-11-19 01:33:19 |
| claude-opus-5 | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-bd3974dc-1795262741989 | 2026-11-21 20:05:41 |
| claude-opus-5-5 | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-86b88728-1795262743151 | 2026-11-21 20:05:43 |
| deepseek-v4-1-flash | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-bb75ebe0-1795262744661 | 2026-11-21 20:05:44 |
| gemini-3-8-flash | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-2211ec78-1795262755137 | 2026-11-21 20:05:55 |
| gemini-4-argon | catalog-only catalog ba2b058babe861d5 | folkbench-connectivity-v1 | rel-1593576e-1795197912236 | 2026-11-21 02:05:12 |
| glm-5-3 | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-d0ab6660-1795262745816 | 2026-11-21 20:05:45 |
| gpt-5-6-terra | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-21fba509-1795262747764 | 2026-11-21 20:05:47 |
| gpt-6-1-sol | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-0619cc9b-1795262751270 | 2026-11-21 20:05:51 |
| gpt-6-astra | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-77658e54-1795262748971 | 2026-11-21 20:05:48 |
| grok-4-6 | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-aed6ecb4-1795262757974 | 2026-11-21 20:05:57 |
| grok-4-7 | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-437141ee-1795262760132 | 2026-11-21 20:06:00 |
| qwen3-8-max | 2026-11-20 21:00:00/2026-11-21 21:00:00(北京时间) | folkbench-connectivity-v1 | rel-34b5b046-1795262761750 | 2026-11-21 20:06:01 |
</details>
**探针汇总:快照收录 263 条探针结果,覆盖 14 家站点。** 汇总一致率:22503/25838(87.1%)。[查看公开结果 →](https://folkbench.com/arena?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-integrity)
[下载对应日期的公开快照](https://github.com/Folkbench/folkbench-data/tree/main/daily/2026/11/21) · [方法说明](https://folkbench.com/methodology?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-method) · [提交纠错](https://folkbench.com/corrections?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=daily-board-corrections)
<!-- FOLKBENCH-BOARD:END -->
---
## Folkbench
Folkbench 把 AI 中转站的公开目录与榜单快照放在一起,让你按模型比较可用率、稳定性、响应速度和成本。选好线路后,用 **Folkbench Switch** 在本地管理 Claude Code、Codex、Gemini CLI 等工具的连接配置。
<table width="100%">
<tr>
<td width="33%" align="center" valign="top"><a href="https://folkbench.com/rankings?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-rankings-4"><img src="docs/images/readme-zh/rankings.webp" width="100%" alt="公开榜单:比较可用率、稳定性、响应速度和成本"></a></td>
<td width="33%" align="center" valign="top"><a href="https://folkbench.com/arena?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-arena-5"><img src="docs/images/readme-zh/integrity.webp" width="100%" alt="降智监控:查看探针一致率、Token 与模型指纹"></a></td>
<td width="33%" align="center" valign="top"><a href="https://folkbench.com/stations?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-stations-6"><img src="docs/images/readme-zh/compare.webp" width="100%" alt="中转站目录:查看公开站点与档案"></a></td>
</tr>
</table>
## 功能入口
| 你关心什么 | 从这里开始 |
| :--- | :--- |
| 哪些线路更常能用? | [可用率榜](https://folkbench.com/rankings?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-rankings-7) |
| 长时间任务的表现怎样? | [稳定性榜](https://folkbench.com/rankings/stability?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-rankings-stability-8) |
| 哪条线路响应更快? | [响应速度榜](https://folkbench.com/rankings/latency?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-rankings-latency-9) |
| 一次编程会话大约花多少? | [会话成本榜](https://folkbench.com/rankings/cost?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-rankings-cost-10) |
| 不同中转站的模型表现是否一致? | [降智监控](https://folkbench.com/arena?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-arena-11) |
| 查找已收录中转站? | [中转站目录](https://folkbench.com/stations?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-stations-12) |
### 降智监控
对同一模型、不同线路查看身份探针题的一致率、输入 Token、思考 Token 和模型指纹;有数据时也可查看缓存检测信息。一次回答或一个百分比不能单独证明模型被替换,结果需要结合检测条件与样本理解。
[打开监控 →](https://folkbench.com/arena?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-arena-13) · [阅读评测方法 →](https://folkbench.com/blog/did-the-model-get-dumber?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-blog-did-the-model-get-dumber-14)
<details open>
<summary><b>Folkbench Rankings</b></summary>
<br>
<img src="docs/images/readme-zh/board.webp" width="960" alt="Folkbench Rankings 榜单界面示意,非实时测量结果。">
示例数据;实际结果见线上榜单。
</details>
<br>
## Folkbench Switch
**把选好的线路,接入每天使用的工具。**
Switch 是基于 Tauri + Rust 的开源桌面客户端。保存服务商的 Base URL 与 API Key,预览配置差异,再应用到目标工具;需要时切换线路或恢复原有配置。
<img src="docs/images/readme-zh/workflow.webp" width="960" alt="使用流程:在网站看公开快照,比较公开信息,再用 Switch 预览并应用本地配置。">
### 下载安装
<p align="center"><b><a href="https://github.com/Folkbench/folkbench/releases">下载 Switch Beta →</a></b><br><sub>macOS · Apple Silicon / Intel | Windows · x64</sub></p>
macOS 安装包为 `.dmg`,Windows 安装包为 `.exe`。
安装包与 SHA256 校验文件见 [Releases](https://github.com/Folkbench/folkbench/releases)。请按具体版本的发布说明选择安装包。
<img src="docs/images/readme-zh/switch.webp" width="960" alt="Folkbench Switch 客户端界面示意:白底、轻量界面与紫色强调的产品展示图。">
### 功能
| 能力 | 使用方式 |
| :--- | :--- |
| **服务商切换** | 保存多家服务的连接参数,将选择应用到目标工具。 |
| **配置差异预览** | 写入前检查变化,保留其他自定义设置;具体恢复能力以工具适配为准。 |
| **Codex 凭据保护** | 切至第三方服务时暂存官方 OAuth 凭据,切回时恢复。 |
| **组合档案** | 把各工具当前连接的线路保存为命名组合,一键整组切换。 |
| **托盘快速切换** | 在系统托盘切换组合或单个工具的服务,无需打开主窗口。 |
| **本机隐私脱敏** | 可选开启。请求经本机代理发出前,把 API Key、邮箱、公网 IP 与自定义敏感词替换为占位符,并在响应中还原;退出 Switch 时恢复直连。 |
| **公开线路发现** | 在客户端查看公开目录中的线路、可用率、状态窗口与价格。 |
| **本地用量估算** | 按会话查看 Token 消耗,并以参考单价估算成本。 |
Switch 在本地管理工具配置。公开目录、模型查询和服务商请求等联网功能仍需要网络;网站上的降智监控与缓存补偿信息不等于客户端已集成的功能。
<details open>
<summary><b>查看支持的工具与配置路径</b></summary>
| 工具 | 配置文件 / 路径 | 说明 |
| :--- | :--- | :--- |
| **Claude Code** | `~/.claude/settings.json` | 自动写入 Base URL,智能处理 Token 与 API Key 凭据类型 |
| **Codex** | `~/.codex/config.toml` | 写入自定义 provider,暂存官方凭据,切回自动恢复 |
| **Gemini CLI** | `~/.gemini/.env` & `settings.json` | 同步写入环境变量与验证方式 |
| **Aider** | `~/.aider.conf.yml` | 写入模型与 API 端点参数 |
| **OpenCode** | `~/.config/opencode/` | 快速注入服务映射与认证 Key |
| **Kimi CLI** | `~/.kimi-code` / `~/.kimi` | 自动兼容新老版本目录 |
| **DeepSeek Harness** | `~/.dsh/` | 适配官方 CLI 环境 |
| **其他** | MiniMax Code、Qwen Code、Grok Build、Hermes、Pi 等 | 持续维护适配中 |
</details>
---
## 社区
- **反馈问题**:[提交 Issue](https://github.com/Folkbench/folkbench/issues),附上复现步骤与版本信息;不要粘贴 API Key 或包含凭据的配置。
- **一起讨论**:[Telegram](https://t.me/+C_E8eL1ZevM4MTY9) · [QQ 群](https://qm.qq.com/q/71NzSTrPVK)(203286875)。
- **阅读与投稿**:[Folkbench 博客](https://folkbench.com/blog?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-blog-15),包含评测方法、模型资料和读者投稿,现有内容覆盖 12 种语言。
- **了解客户端**:[Switch 开发说明](apps/switch/README.md)。
## 引用与徽章
每日归档见 [Folkbench Data](https://github.com/Folkbench/folkbench-data),引用时请注明 Folkbench、归档日期和快照版本。数据许可见该仓库;公开快照不包含原始运行记录,不承诺可以复算。
快照徽章接口:`/badge/{station_slug}.svg?model={model_id}&metric=availability`。嵌入时将徽章链接到对应的主站档案页,并使用 `utm_source=github&utm_medium=badge`。
## 相关链接
- [中转站推荐清单(GitHub)](https://github.com/magnus-minor/relay-api-recommend)
- [中转站实时检测(用自己的 Key)](https://folkbench.com/evaluations/api-quality?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=links-check)
- [Claude 中转站测评](https://folkbench.com/claude-zhongzhuanzhan-ceping?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=links-claude-ceping)
- [GPT / Codex 中转站测评](https://folkbench.com/gpt-codex-zhongzhuanzhan-ceping?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=links-codex-ceping)
- [中转站对比](https://folkbench.com/stations/compare?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=links-compare)
- 公开快照子站(只读 [folkbench-data](https://github.com/Folkbench/folkbench-data) 归档;排序按站点规则,含合作档位与运营加权,部分数据为展示样本):
- [AI 中转站排行](https://folkbench.github.io/relay-rank/):按模型的每日公开快照与历史名次
- [中转站检测周报](https://folkbench.github.io/integrity-weekly/):模型身份探针的每周汇总
- [AI 中转站对比](https://ai-relay-bench.github.io/):同一模型在各中转站的并排对比
- [AI API 价格对比](https://aiapi-rank.github.io/):中转站公开标价与官方参考价
- [AI 中转站大全](https://ai-zhongzhuan.github.io/):站点目录、公开资料与新手指南
- [模型检测快照](https://ai-model-check.github.io/):每日模型身份探针汇总
## 许可证
`apps/switch/` 下的客户端代码以 [Apache-2.0](apps/switch/LICENSE) 开源。博客文章、图片与品牌素材的许可范围见 [LICENSE.md](LICENSE.md)。
<p align="center"><br><b>Folkbench</b><br><sub>在真实场景中衡量 AI 的价值。</sub><br><br><a href="https://folkbench.com/?utm_source=github&utm_medium=readme&utm_campaign=main-repo&utm_content=entry-home-22">访问网站 ↗</a> · <a href="https://github.com/Folkbench/folkbench/releases">下载 Switch ↗</a></p>
Installazione
Installa la skill da https://github.com/Folkbench/folkbenchQuesta skill non è pubblicata sul registro ClawHub. Incolla il prompt nel tuo agente di coding — Claude Code, Cursor, Codex — e installerà la skill direttamente dal repository sorgente.
Apri il repository