Anthropic's Real Competitor Is AWS, Not OpenAI
TL;DR: Foundation models are converging. The top 10 on Chatbot Arena are within 44 ELO points. API prices dropped 80% in a year. Anthropic’s real play is becoming the next AWS, not winning the next benchmark.
Everyone talks about the model race. Claude vs GPT vs Gemini. Who’s smarter, who’s faster, who writes better code.
Try switching models right now. The difference is hard to feel. The race is still going, but the runners are getting harder to tell apart with the naked eye.
The X + Y + Z Framework
Split the AI stack into three layers:
X — Base capability. Pre-training, post-training, raw model intelligence. Where billions get burned, and where the outputs are looking increasingly similar across providers.
Y — The harness layer. Orchestration, guardrails, managed agents, eval frameworks. Everything that takes a fixed X and makes its outputs stable, predictable, and deployable. Claude Managed Agents, agent control planes, and embedded orchestration logic live here.
Z — The customer relationship. Domain expertise, implementation, the actual business contract. This is where revenue lives.
X is converging, Y is the real battleground, Z is where startups should be.
Model Capability Isn’t a Moat Anymore
The Chatbot Arena leaderboard has the top 10 models within 44 ELO points of each other. Claude leads coding, Gemini leads reasoning, GPT leads creative writing. Nobody dominates across the board, and “best model” depends on what you’re doing this afternoon.
On the pricing side it’s even more obvious. API costs fell 80% in one year, with GPT-4o input going from $5 to $2.50/MTok and the cheapest models hitting $0.10. You only see that kind of price war when sellers can’t differentiate.
The recent releases feel the same way. 4.7, 5.5, 4.8, each a bit better than the last, nobody pulling away. “GPT-4 moment” leaps feel further and further apart.
Third-Party Orchestration Has No Future
Why did “harness” as a concept come from model providers, not from startups?
My guess is something like this. Y requires deep model knowledge. Providers see things external developers probably never will: failure modes from post-training, patterns from billions of API calls, what actually works versus what demos well. They’re doing continuous post-training, and best practices grow out of internal observation. Not something you can reverse-engineer from the outside.
Third-party orchestration frameworks (LangChain, CrewAI) are playing catch-up by definition. They can only observe inputs and outputs while the model provider sees the middle. Add the funding gap on top: these startups are competing against providers who can ship orchestration as a free platform feature. How do you fight that?
Y mostly belongs to model providers. Third-party Y tools have a window, but probably not a moat.
Anthropic = AWS, Not OpenAI
If you squint, the cloud playbook is repeating itself.
AWS doesn’t make money selling raw compute. It makes money on Lambda, SageMaker, managed databases. The orchestration layer that turns compute into something deployable is where the margin lives. Underlying compute is a cost center; the services on top are the profit center.
Anthropic is running the same play. Raw model inference is the cost center. Claude Managed Agents, the control plane, enterprise SLAs, that’s the profit center. They even shipped 10 finance agent templates (KYC, earnings review, month-end close). Not to compete with finance startups, but to show enterprises what the platform can do.
The right competitive comparison isn’t Anthropic vs OpenAI on benchmarks. It’s Anthropic vs AWS on who owns the managed environment enterprises build on.
Where Startups Should Be
So where do startups go? Z, the customer layer.
“Won’t Anthropic eat Z too?” Probably not, for the same reason AWS doesn’t do consulting.
Getting prompts, fine-tuning, and workflows right for a domain means living in that domain. You need lawyers who’ve done KYC for ten years, not a model provider shipping templates. Hard to imagine Anthropic simultaneously going deep into legal, healthcare, finance, HR, logistics, and manufacturing. That’s an organizational constraint, not a technical one.
What does a safe Z look like? Two properties:
- Domain immersion — years of industry soaking that a platform provider won’t invest the time for
- Organizational complexity — multi-person, cross-system, process-heavy. Not a one-person prompt job.
One-person use cases (marketing copy, quick summaries) will probably get eaten. But enterprise-scale domain applications are a lot harder to swallow.
Related:
简单说： 模型能力是张入场券，做不到差异化。钱不在推理层，在上面的编排层。Anthropic 的真正路线是成为下一个 AWS。
模型之间的竞赛是谁聪明，谁快，谁代码写得好。
但是你是不是觉得头部的这几家换来换去，体感差别真没多大。比赛还在进行中，但选手之间已经很难用肉眼分出高下了。
X + Y + Z 框架
我把AI stack 分了三层：
X — 基础能力。 Pre-training、post-training、裸模型智力。烧几十亿的地方，各家出来的东西越来越像。
Y — Harness 层。 Orchestration、guardrails、managed agents、eval 框架。在固定的 X 之上让输出稳定、可预测、可部署。Claude Managed Agents、agent control plane、嵌入模型层的编排逻辑都在这一层。
Z — 客户关系。 Domain expertise, business relationship, implementation, 真正的商业合同。Revenue 活在这里。
X 在趋同，Y 是真正的战场，Z 是创业者该待的地方。
模型能力已经不是壁垒
Chatbot Arena 排行榜上前十名之间只差 44 ELO。Claude 领 coding，Gemini 领 reasoning，GPT 领创意写作，没谁全面碾压。”最好的模型”看你下午干嘛。
API 价格那边更明显，一年降了 80%，GPT-4o input 从 $5 降到 $2.50/MTok，最便宜的到了 $0.10。卖方没法差异化的时候才会打这种价格战。
最近几个版本也是这个感觉。4.7、5.5、4.8，每个比上一个好一点点，但没谁拉开距离。”GPT-4 时刻”那种代际跃升，感觉离我们越来越远了。
第三方 orchestration 没有明天
一个有意思的观察：为什么 “harness” 这个概念是 model provider 提出的，不是创业者？
我猜大概是这样的。Y 需要深度的模型知识，provider 能看到外部开发者大概永远看不到的东西：post-training 中浮现的 failure mode、数十亿 API call 里积累的 pattern、什么真正 work 什么只是 demo 好看。他们在持续 post-training，best practice 是从内部观察里长出来的，外面逆向工程不了。
第三方 orchestration 框架（LangChain、CrewAI）定义上就是在追。它们只能观察 input/output，model provider 看到的是中间过程。再加上这些创业公司在跟能把 orchestration 做成免费 platform feature 的 model provider 竞争，这仗很难打。
Y 大概率属于 model provider。第三方 Y 工具有窗口期，但很可能没有 moat。
Anthropic = AWS
AWS 不靠卖裸算力赚钱，靠的是 Lambda、SageMaker、managed databases 这些东西。编排层把算力变成可部署的产品，利润在那里。底层算力是成本中心，上层服务才是利润中心。
Anthropic 在跑同样的 playbook。裸模型推理是成本中心，Claude Managed Agents、control plane、enterprise SLA 才是利润中心。他们甚至发了 10 个金融 agent 模板（KYC、earnings review、month-end close），不是要跟金融创业公司抢生意，是在告诉企业：看，平台能干这个。
正确的竞争对比不是 Anthropic vs OpenAI 比跑分，是 Anthropic vs AWS 比谁拿下企业的 managed environment。
创业者该待的地方
那创业者去哪？Z，客户层。
“Anthropic 不会吃 Z 吗？” 大概率不会，跟 AWS 不做咨询一个道理。
要把一个 domain 的 prompt、fine-tuning、workflow 做对，你得泡在那个行业里。你需要的是做了十年 KYC 的律师，不是 model provider 发的一套模板。很难想象 Anthropic 同时深入法律、医疗、金融、HR、物流、制造业，这是组织能力的约束，不是技术问题。
安全的 Z 长什么样？两个特征：
- Domain immersion，需要年深日久的行业浸泡，platform provider 不会投入这个时间
- 组织复杂度，多人协作、跨系统、流程重，不是一个人写几句 prompt 就能搞定的
一个人能搞定的轻场景（营销文案、快速摘要）大概率会被吃掉。但 enterprise-scale 的 domain application，要吃下去就难多了。
相关文章：