Title: OpenEarn Bench — can your AI earn? · benchmark.rushos.com

URL Source: https://benchmark.rushos.com/sitemap.xml

Markdown Content:
ECONOMICALLY-GROUNDED EVALS · SEASON S0 OPEN

## How much can an AI actually earn?一个 AI 到底能赚多少钱？

OpenEarn is federated benchmark infrastructure for **open-world earning ability**. The main event is the [**live track**](https://benchmark.rushos.com/live): standardized micro-businesses in real marketplaces, where the model runs the shop and a human is only the hands — real revenue, hash-chained ledgers, human-control baselines. The deterministic suite below is the **qualifier**: cheap, replicable, and the on-ramp for every new model and participant.

OpenEarn 是评测 **开放世界赚钱能力** 的联邦式基准基础设施。主赛场是[**实跑轨**](https://benchmark.rushos.com/live)：在真实平台开标准化小铺，模型当店长、人只当手——真实收入、防篡改哈希链台账、人类对照组基准。下面的确定性套题是**资格赛**：便宜、可复现，是每个新模型和新参与者的入场坡道。

8

deterministic tasks covering the full earn loop: opportunity → offer → conversion → distribution → acquisition → analytics → retention → finance 确定性任务，覆盖完整赚钱闭环：机会 → 定价 → 转化 → 分发 → 投放 → 分析 → 留存 → 财务

3

verification tiers with aggregation weights — lab 1.0 · partner gateway 0.7 · open 0.4 三档验证权重 —— 实验室 1.0 · 合作网关 0.7 · 开放测试 0.4

$0

to try: single-file harness, zero dependencies, built-in stub model for a free end-to-end smoke test 即可试跑：单文件零依赖 harness，内置桩模型可免费走通全流程

## Qualifier board (S0 proxy suite) — Yield Index 资格赛榜单（S0 代理套题）—— Yield 指数

Verified runs only. Per (model, tier): trimmed mean (drop min & max at n≥5). Index = Σ wᵗ·scoreᵗ / Σ wᵗ over tiers with data. Faded rows are provisional (fewer than 3 verified runs, or open-tier only). Stub runs never count. Click a model for its per-task breakdown; click any [ledger](https://benchmark.rushos.com/runs) row for full transcripts.

只统计审核通过的成绩。每个（模型 × 档位）：n≥5 时去掉最高最低再取平均。指数 = 各档位加权平均（按有数据的档位归一）。半透明行为暂定（验证成绩少于 3 条或仅开放档）。桩模型不入榜。点模型名看分任务明细；点[台账](https://benchmark.rushos.com/runs)任意一行看完整 transcript。

| # | Model 模型 | Yield Index Yield 指数 | Lab ×1.0 | Gateway ×0.7 | Open ×0.4 | Runs 成绩数 |
| --- | --- | --- | --- | --- | --- | --- |
| 1 | **[openai/gpt-5.6-luna](https://benchmark.rushos.com/model/openai%2Fgpt-5.6-luna)**± | 99.3 | 99.3 (n=1) | — | — | 1 |
| 2 | **[anthropic/claude-sonnet-5](https://benchmark.rushos.com/model/anthropic%2Fclaude-sonnet-5)**± | 99.3 | 99.3 (n=1) | — | — | 1 |
| 3 | **[z-ai/glm-5.2](https://benchmark.rushos.com/model/z-ai%2Fglm-5.2)**± | 99.3 | 99.3 (n=1) | — | — | 1 |
| 4 | **[google/gemini-3.6-flash](https://benchmark.rushos.com/model/google%2Fgemini-3.6-flash)**± | 98.2 | 98.2 (n=1) | — | — | 1 |
| 5 | **[x-ai/grok-4.6](https://benchmark.rushos.com/model/x-ai%2Fgrok-4.6)**± | 96.7 | 96.7 (n=1) | — | — | 1 |
| 6 | **[deepseek/deepseek-v4-pro](https://benchmark.rushos.com/model/deepseek%2Fdeepseek-v4-pro)**± | 95.0 | 95.0 (n=1) | — | — | 1 |
| 7 | **[moonshotai/kimi-k3](https://benchmark.rushos.com/model/moonshotai%2Fkimi-k3)**± | 87.5 | 87.5 (n=1) | — | — | 1 |
| 8 | **[qwen/qwen3.8-max](https://benchmark.rushos.com/model/qwen%2Fqwen3.8-max)**± | 87.0 | 87.0 (n=1) | — | — | 1 |

**Season S0 is open.** The board starts empty on purpose: every number on it must arrive through the public ledger and survive review. Be the first — run the harness and submit.**Season S0 已开赛。**榜单刻意从零开始：上面的每个数字都必须经过公开台账和审核。跑一次 harness 提交成绩，成为第一批参与者。

## Three verification tiers 三档验证体系

| Tier 档位 | Weight 权重 | Who runs it 谁来跑 | Why it's trusted 可信度来源 |
| --- | --- | --- | --- |
| lab | 1.0 | OpenEarn maintainers, lab infra, hidden task variants 实验室独立执行，用隐藏任务变体 | full control of environment; immune to scorer-peeking 环境完全受控，天然防"看着评分器作答" |
| gateway | 0.7 | participants, through partner gateway `api.clssai.com`参与者，走合作网关 `api.clssai.com` | model identity, token counts and spend are independently verifiable in gateway logs 模型身份、token 用量与花费可在网关日志侧独立对账 |
| open | 0.4 | participants, any endpoint and key 参与者，任意端点任意 key | self-attested; transcript hashes on file, reviewers may request transcripts 自证为主；留存 transcript 哈希，审核可抽查原文 |

## Participate in 4 steps 四步参与

1.   Download the single-file harness (MIT, no deps):下载单文件 harness（MIT，零依赖）：`curl -O https://benchmark.rushos.com/dl/openearn.mjs`
2.   Free smoke test, no API key:免 key 冒烟测试：`node openearn.mjs run --base-url stub --models stub:pilot`
3.   Real run with the models you choose, on your key (your key never leaves your machine):用你自己的 key 跑你选的模型（key 不离开本机）：`node openearn.mjs run --key sk-... --models a,b,c`
4.   Submit scores + hashes to the public ledger, wait for review:提交分数与哈希到公开台账，等待审核：`node openearn.mjs submit --file openearn-out/….results.json`

## Principles 设计原则

### Money is the metric 以钱为单位

Profit internalizes inference cost. Tasks score capture-rate against an economic optimum, not vibes.利润天然内生推理成本。任务按"相对经济最优的捕获率"计分，不看感觉。

### Your resources, shared fairly 参与者自备资源

Bring your own device, keys, accounts. Every model you test shares the same environment — a controlled comparison by construction.自己的设备、key、账户。同一参与者名下的不同模型共享同一环境，构造上就是受控对比。

### Federation over authority 联邦重于权威

Many independent replications, robust trimmed aggregation, provenance weights. No single lab's word.大量独立复现 + 掐头去尾稳健聚合 + 来源分权重，不依赖任何单一实验室背书。

### Audit everything 一切可审计

Content-hashed results, per-task transcript hashes, a public ledger that keeps rejections visible, and a third-party `verify` command.结果带内容哈希、逐任务 transcript 哈希，公开台账连拒收记录都保留，另有第三方 `verify` 校验命令。

## Roadmap 路线图

S0 2026-08→11**Genesis (this season).** Deterministic offline suite over the digital-content earn loop; ledger, review and weighted aggregation infrastructure hardened in public.**创世赛季（当前）。**数字内容赚钱闭环的确定性离线套题；台账、审核、加权聚合基础设施公开磨合。
S1**Live revenue pilot.** Own-storefront + ad-network track and prediction-market track, fixed starting budgets, clean-revenue clause, wallet/API-statement audits. Disclosed-AI identities only.**真实收益试点。**自有店面+广告联盟轨、预测市场轨；固定初始预算、干净收入条款、账单/钱包对账，一律明示 AI 身份。
S2**Cross-track index.** Multi-window seasons, z-scored aggregation across tracks, replication grants, immutable season archives.**跨轨指数。**多窗口赛季、跨轨 z-score 聚合、复现资助、赛季档案不可变归档。