95%的折扣，从来不是白给的。
Meta为旗下编码与agent模型Muse Spark上线了"贡献者定价"：输入token每百万0.1美元、输出token每百万0.2美元——对照标准价的1.25美元与4.25美元，价差约12倍与21倍，按TechCrunch的算法约等于95%折扣。折扣的条件只有一行：接受这个价格，即同意Meta将你的数据用于模型训练。
定价表里藏着一行小字
这套机制的聪明之处在于"自愿"：没有人被强制收集数据，交换以折扣形式摆在台面上。但对预算敏感的开发者与初创团队，0.1美元的输入价几乎无法拒绝——尤其当隔壁标准价是1.25美元。敏感数据、客户代码、商业机密，每一项都得在"省钱"与"交出去"之间重新称重。TechCrunch还提到一个背景：Meta内部一个员工追踪项目已于6月被暂停——这家公司对"数据来源"的胃口，由来已久。
为什么开发者数据突然值钱了
两个署名观点值得听。普林斯顿计算机科学家Arvind Narayanan的判断是：企业客户宁可买贵的Enterprise版本，也不愿把内部数据交出去——这意味着贡献者定价瞄准的是另一群人：独立开发者、学生、小团队。开发者Mario Zechner则从另一面论证了这门生意的逻辑：Claude Code等编码工具的成功，已经证明真实开发者数据驱动强化学习的价值。换句话说，你的终端命令与代码上下文，在2026年的估值体系里，已是接近燃料级的资产。
一场自愿的交易，一个岔路口
"以折扣换数据"最深远的影响，是把数据获取从灰色地带拉进了明码标价：行业第一次有了公开价签——你的代码，值95%的账单减免。它比爬虫体面，比授权付费便宜，大概率会被同行抄走。但对个体开发者，岔路口上真正的问题是：当模型公司愿意付"负价格"买你的数据时，你的数据到底值多少？95%折扣给出的，只是下限。
明码标价不是终点，是讨价还价的开始。
明天见。
A 95% discount is never free.
Meta has rolled out "contributor pricing" for its coding and agent model Muse Spark: $0.10 per million input tokens and $0.20 per million output tokens — against list prices of $1.25 and $4.25, gaps of roughly 12x and 21x, or about a 95% discount by TechCrunch's math. The condition fits on one line: accept this price, and you consent to Meta training on your data.
The fine print on the price sheet
The mechanism's cleverness lies in "voluntary": nobody is forcibly collected from; the trade sits openly on the table as a discount. But for budget-sensitive developers and startups, $0.10 input pricing is nearly irresistible — especially with $1.25 on the next shelf. Sensitive data, client code, trade secrets — each now needs re-weighing between "saving money" and "handing it over." TechCrunch adds context: an internal Meta employee-tracking project was paused in June — this company's appetite for "data sources" is of long standing.
Why developer data suddenly became precious
Two named takes are worth hearing. Princeton computer scientist Arvind Narayanan's judgment: enterprise customers would rather buy the pricier Enterprise tier than hand over internal data — which means contributor pricing targets a different crowd: independent developers, students, small teams. Developer Mario Zechner argues the other side of the business: the success of coding tools like Claude Code has already proven the value of real developer data for driving reinforcement learning. Your terminal commands and code context, in 2026's valuation system, are fuel-grade assets.
A voluntary trade at a fork
The deepest impact of "discount-for-data" is dragging data acquisition out of the grey zone and into listed prices: the industry's first public tag — your code, worth a 95% reduction on the bill. It is more decent than scraping and cheaper than licensing, and competitors will surely copy it. But for the individual developer, the real question at the fork: when a model company is willing to pay a negative price for your data, what is your data actually worth? The 95% discount marks only the floor.
Listed prices aren't the end of negotiation — they're the start.
See you tomorrow.
折扣是明码的，数据是暗码的——现在两者印在同一张价签上。
—— Dawn Vision编辑部
The discount is priced in the open; the data isn't — now both are printed on the same price tag.
— The Dawn Vision Editorial Desk
Meta为Muse Spark（coding/agents模型）推出贡献者定价 · $0.10/M输入（标准$1.25）、$0.20/M输出（标准$4.25），价差约12倍（输入）/21倍（输出），TechCrunch计算口径约95%折扣 · 条款：接受定价即同意数据用于模型训练 · 背景：Meta内部员工追踪项目6月被暂停（TechCrunch口径） · 观点：Arvind Narayanan（普林斯顿）——企业宁买Enterprise不交数据；Mario Zechner——Claude Code成功证明真实开发者数据驱动RL的价值 · 机制为自愿交换，非强制收集
Meta launches contributor pricing for Muse Spark (coding/agents model) · $0.10/M input (list $1.25), $0.20/M output (list $4.25); gaps ~12x input / 21x output; ~95% discount by TechCrunch's math · Terms: accepting the price means consenting to training on your data · Context: internal Meta employee-tracking project paused in June (TechCrunch) · Takes: Arvind Narayanan (Princeton) — enterprises prefer pricey Enterprise tiers over surrendering data; Mario Zechner — Claude Code's success proves real developer data drives RL · Mechanism is a voluntary trade, not forced collection
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理当日采集信号并核对核心来源后生成，经编辑部人工审核。素材来源：TechCrunch原文（定价数字、条款机制、Narayanan与Zechner观点、员工追踪项目背景）。95%折扣为TechCrunch计算口径，正文同时给出两档价格原文；机制为'可选择换折扣'，正文未作'强制收集'表述。
This article was generated by the Dawn Vision cognitive engine processing collected daily signals against the core source, followed by human editorial review. Source: TechCrunch original (pricing figures, terms mechanism, Narayanan and Zechner quotes, employee-tracking context). The 95% discount is TechCrunch's calculated framing; both list prices are given in text. The mechanism is an opt-in trade; no 'forced collection' claim is made.
