Databricks AI Index:定价看任务成本、分层防高耗、多模型路由
5 Interesting Learnings from the Databricks AI Index: Open-Weight Jumps to 31% of Enterprise Tokens, 72% of Companies Now Use Multiple Model Providers, and a 281x Gap Between Median and Power Users
直接给出 AI 产品定价的核心判据(按任务而非 token)和用户分层策略,数据详实,创业者可立即用于调整定价模型和架构选型。
Most model share data we get comes from consumer apps, developer routers, or surveys. The Databricks AI Index uses a different source: actual token traffic from more than 10,000 companies that route their AI through Databricks’ Unity Gateway. That includes Fortune 500 and Global 2000 companies, plus traffic Databricks sends out to Anthropic and OpenAI.
我们获取的大多数模型共享数据来自消费者应用、开发者路由器或调查。Databricks AI Index 使用了不同的来源:超过 10,000 家通过 Databricks Unity Gateway 路由其 AI 流量的公司的实际令牌流量。其中包括财富 500 强和全球 2000 强公司,以及 Databricks 发送给 Anthropic 和 OpenAI 的流量。
One caveat before the data. Databricks hosts open-weight models and has a commercial interest in companies using them, so its customers probably lean further toward open-weight than the average enterprise. Even with that discount, the speed of what changed here is hard to ignore.
在展示数据之前有一个注意事项。Databricks 托管开源权重模型,并且有商业利益促使公司使用这些模型,因此其客户可能比一般企业更倾向于开源权重。即使考虑到这一折扣,这里发生变化的速度也是不容忽视的。
1. Open-Weight Went From 7.3% to 30.9% of Enterprise Tokens in About Four Months
1. 在大约四个月内,开源权重占企业令牌的比例从 7.3% 上升到 30.9%
Open-weight share of enterprise token traffic: 7.3% in July, 30.9% in November.
企业令牌流量中开源权重的占比:7 月为 7.3%,11 月为 30.9%。
Patrick Wendell, a Databricks co-founder, said why in his launch thread. Anecdotally, companies use them mostly to cut costs on high-volume work. Enterprises are not swapping open-weight in for Claude on their hardest problems. They’re moving the bulk work there: classification, extraction, summarization, the stuff that runs a million times a day.
Databricks 联合创始人 Patrick Wendell 在其发布线程中解释了原因。据 anecdotal 证据显示,公司主要使用它们来降低高容量工作的成本。企业在处理最困难的问题时并没有用开源权重替换 Claude。他们是将大量工作转移到了那里:分类、提取、摘要,即每天运行百万次的那些任务。
Two details stand out:
有两个细节引人注目:
- The open-weight leaders are mostly Chinese labs. The index lists Zhipu (GLM), DeepSeek, and Moonshot (Kimi) as the open-weight providers. These are large enterprises running Chinese-origin weights on US infrastructure, at volume. A year ago most enterprise security teams would have said no.
- One model release can move the whole category. GLM 5.2 accounts for about 40% of open-weight tokens on the gateway, one month after its mid-July availability.
- 开源权重的领导者主要是中国实验室。该指数列出智谱(GLM)、DeepSeek 和 Moonshot(Kimi)作为开源权重提供商。这些是在美国基础设施上大规模运行源自中国的权重的大型企业。一年前,大多数企业安全团队会说不。
- 一个模型的发布可以推动整个类别的变化。GLM 5.2 在 7 月中旬发布后一个月,占据了网关上约 40% 的开源权重令牌。
What this means if you’re building: If your product makes a model call on every record, every ticket, or every email, a third of your enterprise buyers’ traffic now runs on models that cost a fraction of frontier pricing. Your competitors will rebuild their COGS on open-weight. If you still route everything to a frontier model, you’ll lose on gross margin before you lose on product.
如果你正在构建产品,这意味着什么:如果你的产品在每条记录、每个工单或每封邮件上都调用模型,那么现在你三分之一企业买家的流量都在运行成本仅为前沿定价一小部分的模型上。你的竞争对手将基于开源权重重建他们的 COGS。如果你仍然将所有内容路由到前沿模型,那么在产品优势丧失之前,你就会失去毛利率。
2. Single-Provider Companies Fell From 64% to 28%
2. 单一提供商的公司比例从 64% 下降到 28%
Enterprises on Unity Gateway relying on one model provider: about 64% in July, falling to 28% by November 2. That happened in one quarter.
依赖单一模型提供商的 Unity Gateway 上的企业:7 月约为 64%,到 11 月 2 日降至 28%。这发生在一个季度内。
Databricks’ earlier data shows the same trend at the model level. Average unique models per customer rose from 4.2 to 7.1 since January. Roughly 40% of those models were open-weight.
Databricks 的早期数据显示了模型层面的相同趋势。自 1 月以来,每位客户的平均唯一模型数量从 4.2 增加到 7.1。其中大约 40% 的模型是开源权重的。
Enterprises now treat models the way they treat cloud regions: several in use, routed by workload, swapped out when something better ships. It matters for two reasons:
企业现在对待模型的方式,就像对待云区域一样:同时使用多个,按工作负载路由,当有更好的版本推出时便进行替换。这之所以重要,有两个原因:
- “We’re built on GPT” or “we’re built on Claude” isn’t a moat, and enterprise buyers don’t want to hear it. They want to know you can use whichever model they’ve already approved, including the open-weight ones running inside their own perimeter.
- The release pace makes lock-in impossible anyway. Since May 2026, a new flagship model launched roughly every week, or every 3.7 days when regional labs are included.
- “我们基于 GPT 构建”或“我们基于 Claude 构建”并非护城河,企业买家也不想听这些。他们想知道的是,你能使用他们已经批准的任意模型,包括运行在他们自身安全边界内的开源权重模型。
- 发布节奏使得锁定(lock-in)无论如何都成为不可能。自 2026 年 5 月以来,大约每周就会推出一款新的旗舰模型;如果算上区域性实验室,则平均每 3.7 天就有一款新旗舰模型发布。
3. Anthropic Still Owns the Proprietary Side, and Opus 5.5 Took Over in Six Weeks
3. Anthropic 仍占据专有模型领域的主导地位,Opus 5.5 在六周内接管了市场
Open-weight is growing, but Anthropic’s lead within proprietary models grew too. Anthropic’s overall share rose from 57.6% to 68.8% over the period covered, which the article ties to Opus 5.5’s launch.
开源权重模型的市场份额在增长,但 Anthropic 在专有模型领域的领先地位也在扩大。在文章涵盖的期间内,Anthropic 的整体份额从 57.6% 上升至 68.8%,文章将此归因于 Opus 5.5 的发布。
How fast it happened is the real point. Claude Opus 5.5 went from 7.2% to 37.9% of proprietary-model tokens in six weeks. Enterprises are supposed to be slow. These ones moved a third of their proprietary traffic to a new model in a month and a half, because a gateway turns switching into a config change.
真正关键的是发生速度有多快。Claude Opus 5.5 在六周内占据了专有模型 token 总量的 7.2% 至 37.9%。企业通常以行动缓慢著称,但这些企业却在短短一个半月内将三分之一的专有流量转移到了新模型上,因为网关将切换过程简化为一次配置更改。
The flip side matters just as much. The best model on Databricks’ coding benchmark is Claude Fable 5.1, yet Fable 5.1’s share of proprietary tokens fell from 5.5% to 0.6% over six weeks. The most capable model is not the most used. Once a model is good enough at a lower price, enterprises move to it.
另一方面同样重要。Databricks 编程基准测试中表现最好的模型是 Claude Fable 5.1,然而 Fable 5.1 在专有模型 token 中的份额在六周内从 5.5% 下降至 0.6%。能力最强的模型并非使用最广泛的模型。一旦某模型在更低的价格下达到足够好的水平,企业就会转向它。
Databricks did the same thing internally. They kept Opus 4.8 rather than upgrading to Opus 5, because the upgrade wasn’t worth the added cost for its workloads.
Databricks 内部也采取了同样的做法。他们保留了 Opus 4.8,而没有升级到 Opus 5,因为对于其工作负载而言,升级带来的成本增加并不值得。
4. Gemini 3.8 Flash Had a Low Token Price and Cost $11.97 Per Task
4. Gemini 3.8 Flash 具有较低的 token 价格,每任务成本为 11.97 美元
This is the most useful section for anyone pricing an AI product. Databricks built a coding benchmark from real merged pull requests in its multi-million-line codebase and measured cost per task, not cost per token.
这对于任何为 AI 产品定价的人来说都是最有用的部分。Databricks 在其数百万行代码库的真实合并拉取请求基础上构建了一个编程基准测试,并衡量每个任务的成本,而非每个 token 的成本。
- Best quality: Claude Fable 5.1 at 96.0% task success, $7.47 per task. Muse Wonder: 94.9% at $1.53 per task. That extra 1.1 points costs almost 5x as much.
- GPT-6 Luna and Claude Opus 5.5 both hit 89.1%. Luna averaged $0.05 per task versus $1.62 for Opus 5.5, about 32x less.
- Gemini 3.8 Flash averaged 140 turns per task and $11.97 per task, one of the most expensive in the benchmark despite a low token price.
- On open-weight, DeepSeek V4.1 and GLM 5.3 are both third, with a 93.1% pass rate, which is ahead of Opus 5.5 on this benchmark.
- 最佳质量:Claude Fable 5.1,任务成功率为 96.0%,每任务成本 7.47 美元。Muse Wonder:任务成功率为 94.9%,每任务成本 1.53 美元。这额外的 1.1 个百分点的成本几乎高出近 5 倍。
- GPT-6 Luna 和 Claude Opus 5.5 均达到 89.1% 的任务成功率。Luna 平均每任务成本为 0.05 美元,而 Opus 5.5 为 1.62 美元,前者约为后者的 32 分之一。
- Gemini 3.8 Flash 平均每任务进行 140 轮对话,每任务成本为 11.97 美元,尽管其 token 单价较低,但在该基准测试中属于成本最高的模型之一。
- 在开源权重模型方面,DeepSeek V4.1 和 GLM 5.3 均位列第三,通过率为 93.1%,在该基准测试中优于 Opus 5.5。
The harness matters almost as much as the model. In a separate test, the same model scored 61.7% in one harness, 48.1% in another, and 38.4% bare.
线束(harness)的重要性几乎与模型本身相当。在另一项测试中,同一模型在线束A中得分61.7%,在线束B中得分48.1%,裸跑时得分38.4%。
The decision for founders: Stop judging models by price per million tokens. A cheap model that takes 140 turns to finish costs more than an expensive one that finishes in 10. Measure cost per completed task on your own workload. Same score, 32x cost difference. That’s your gross margin.
创始人的决策:停止按每百万token的价格来评判模型。一个便宜但需要140轮对话才能完成任务的模型,其成本高于一个昂贵但仅需10轮即可完成的模型。请根据您自身的工作负载,衡量每个已完成任务的成本。得分相同的情况下,成本差异可达32倍。这就是您的毛利空间。
5. The Median User Spends $2.49 a Week. The P99 User Spends $699.74.
5. 中位数用户每周花费2.49美元。第99百分位(P99)用户花费699.74美元。
This is the number I’d put in front of every B2B CEO still pricing AI per seat.
这是我想要放在那些仍按席位定价AI的B2B首席执行官面前的数字。
Median: $2.49. It more than doubled from $1.10 over six weeks. P90: no value given. The article says it stayed “relatively flat.” P99: $699.74, about 281x the median.
中位数:2.49美元。六周内从1.10美元以上翻倍。P90:未提供数值。文章称其保持“相对平稳”。P99:699.74美元,约为中位数的281倍。
$700 a week works out to about $36,000 a year in list-price tokens for one person. That’s more than most companies pay per seat for any software they own. And the median user doubled their spend in six weeks.
每周700美元相当于每人每年清单价格下的token费用约为36,000美元。这比大多数公司为其拥有的任何软件支付的单席位费用都要高。而且,中位数用户在六周内将其支出翻了一番。
It means three things:
这意味着三点:
- Flat per-seat pricing breaks at the top of the distribution. Your heaviest 1% of users can wipe out the margin from everyone else. You need usage caps, credits, or tiers that track consumption.
- The budget is real, and it isn’t software budget. I’ve said I won’t invest in an agentic B2B company unless it unlocks budget beyond software budget. Paying $36K a year for one power user’s tokens is the clearest sign yet that this budget exists. It’s being spent on headcount-level output, not on seats.
- CFOs are about to notice. Databricks titled one of its own posts “The AI Bill is coming.” Expect enterprise buyers to ask you for per-user budgets, routing, and visibility into spend before they’ll sign.
- 固定的单席位定价在分布顶端失效。您最重的1%用户可能会抹去其他所有用户的利润。您需要设置使用上限、信用额度或跟踪消耗的层级。
- 预算是真实存在的,且它并非软件预算。我曾表示,除非一家代理型(agentic)B2B公司能解锁超出软件预算之外的资金,否则我不会投资。为单个重度用户每年支付36,000美元的token费用,是目前证明这一预算存在的最清晰信号。这笔钱花在了人力级别的产出上,而非席位上。
- 首席财务官(CFOs)即将注意到这一点。Databricks在其一篇帖子中题为“AI账单即将到来”。预计企业买家在签署合同前,会要求您提供按用户划分的预算、路由机制以及支出可见性。
A Few More Interesting Learnings
更多有趣的发现
- Better models can cost less. Inside Databricks, Opus 5.5 delivered improved quality at 29% lower cost per session than its predecessor. Upgrading isn’t always a price increase.
- Frontier guardrails are pushing some work to open-weight. Hosted frontier models refused Hugging Face responders’ analysis of attack logs and exploit payloads. The team used GLM 5.2 instead. Security and other edge-case workloads will move to models you can run yourself.
- Data retention is now a selling point. Databricks says frontier providers have departed from zero-data-retention, including a 30-day minimum retention tied to Anthropic’s frontier tier. If you sell to regulated industries, expect retention terms to become part of model selection.
- The #1 model keeps changing. The top intelligence ranking has changed hands 20+ times since May 2023. Build the routing layer now, because the model you depend on today won’t be the leader next quarter.
- Wasted agent tokens are a large, hidden cost. One Databricks post describes cutting $1 million a year of wasted agent spend in one hour by fixing bugs in MCP tool servers. If you run agents in production and nobody is tracing tool calls, you’re probably paying for loops.
- 更好的模型可能成本更低。在Databricks内部,Opus 5.5相比其前身,在质量提升的同时,每次会话成本降低了29%。升级并不总是意味着价格上涨。
- 前沿模型的护栏正将部分工作推向开源权重模型。托管的前沿模型拒绝了Hugging Face响应者对攻击日志和利用载荷的分析。团队转而使用了GLM 5.2。安全和边缘案例等工作负载将转向您可自行运行的模型。
- 数据保留现已成为卖点。Databricks指出,前沿提供商已背离零数据保留政策,包括与Anthropic前沿层级挂钩的最低30天保留期。如果您向受监管行业销售,预计保留条款将成为模型选择的一部分。
- 排名第一的模型一直在变。自2023年5月以来,顶级智能排名的宝座已易手20多次。现在就构建路由层,因为你今天依赖的模型下季度可能就不再是领头羊了。
- 浪费的智能体(Agent)Token 是一笔巨大的隐性成本。Databricks 的一篇博文指出,通过修复 MCP 工具服务器中的 bug,仅用一小时就削减了每年100万美元的无效智能体支出。如果你在在生产环境中运行智能体且无人追踪工具调用,那你很可能在为循环调用买单。
Sources:
来源:
- Databricks AI Index
- Introducing the Databricks AI Index (Databricks blog)
- Databricks AI Index
- 发布 Databricks AI Index(Databricks 博客)
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力