10xAI Wiki

15分钟阅读

Anthropic & Claude

Company Overview

Anthropic is an AI safety company founded in 2021 by Dario Amodei, Daniela Amodei, and other former OpenAI researchers. Its mission is to help humanity safely navigate the transition to transformative AI. The company occupies a unique position: it believes AI could be extremely dangerous, yet continues developing it under the reasoning that safety-focused labs at the frontier are better than ceding that ground to less safety-conscious actors.

Key Facts (as of early 2026)

  • Founded: 2021
  • Headquarters: San Francisco
  • Valuation: $380 billion (February 2026) — matching Alibaba’s market cap at the time
  • Cumulative funding: $53.2 billion+ across 9 rounds
  • 2026 Round: $30 billion led by Goldman Sachs Capital, GIC (Singapore sovereign wealth fund); with Microsoft and NVIDIA participation
  • Annual revenue (2025): 14 billion annualized, up 40% from 10B in 2024
  • Revenue split: 80%+ from enterprise customers
  • Profitability: Not yet profitable; breakeven target pushed to 2028
  • IPO: Preparing for 2026 H2; engaged Wilson Sonsini

Competitive Position vs. OpenAI

DimensionAnthropicOpenAI
Valuation$380B$500B
Revenue modelEnterprise subscriptions (80%+)Diversified (consumer + enterprise)
ARPU~$211/MAU~$25/MAU
Core strengthSafety, enterprise, codingConsumer scale, multimodal, brand
Flagship modelClaude Opus 4.6GPT-5.4

Claude Model Family

Claude is Anthropic’s production model line. It is the company’s main commercial product and direct embodiment of its safety mission.

Current Tier Structure

  • Haiku: Fastest, cheapest; for light tasks and high-volume applications
  • Sonnet: Balanced performance and cost; most popular for enterprise workloads
  • Opus: Highest capability; frontier coding and complex reasoning

Key Performance (2025–2026)

  • Claude Sonnet 4.5: 72.7% on SWE-bench Verified — highest published score
  • Claude Opus 4.6: 72.5% SWE-bench; historically dominant in coding (“断崖式领先” — cliff-edge leading); even Google and OpenAI engineers use it internally
  • Pricing: Opus at 15/M input, 75/M output; Sonnet at 3/M input, 15/M output

Model Characteristics

Claude is notable for:

  • Strongest coding performance: The de facto standard for AI-assisted software development
  • Safety and alignment: Most rigorous Constitutional AI implementation
  • Long context: Competitive context windows
  • Enterprise reliability: High uptime, compliance focus

Products & Ecosystem

Claude.ai

Consumer and professional chat interface. Free tier + Pro at 20/month + Max at 100–200/month.

Claude Code

AI coding agent (CLI tool). Understands entire codebases, reads/writes code, runs tests, submits PRs autonomously. As of early 2026, annualized revenue of $2.5 billion — the primary growth driver. Enterprise user count grew 4x from start of 2025.

Key differentiator: Not a “co-pilot” (suggestions you implement), but a “designated driver” — you specify destination, it drives. Capable of handling 18,000-line React components other AI agents couldn’t touch.

Competitors: Cursor, GitHub Copilot, Windsurf — but Claude Code was first to truly autonomous, multi-step code execution.

Model Context Protocol (MCP)

See MCP Protocol page.

Anthropic invented MCP as the standard for AI model tool integration — effectively the “USB-C” of AI agents. It allows any AI agent to connect to external tools (GitHub, Google Drive, Slack, databases) through a standard interface.

Strategic significance: Whoever controls the protocol layer controls the ecosystem. Even OpenAI is adopting MCP compatibility. This is Anthropic’s “TCP/IP” play — more strategic than any single model.

Cowork

AI colleague for non-technical users — handles file organization, document processing, presentations. Born from Claude Code being used for non-coding tasks.

2026 产品分界线（知识工作者 vs 开发者）：

产品面向场景
Claude Code开发者终端写代码
Claude Cowork知识工作者浏览器做文档、跨工具协作

knowledge-work-plugins（2026-06）

Anthropic 官方开源的 20 个岗位插件集（19K+ Star，Apache 2.0）。为 Claude Cowork + Claude Code 提供销售、客服、产品、法务、金融、数据、营销、HR、工程等岗位的专业技能和工具连接。纯 Markdown + JSON，无需写代码即可装载到 Claude。

每个插件连接真实 SaaS（HubSpot、Slack、Jira、Snowflake、Figma 等）。战略意义：Claude Cowork 正式走向企业全函数渗透，与 OpenAI Operator 在企业工作流层面直接竞争。

来源：2026-06-09 Batch

Claude in Chrome

Browser extension that automates web interactions — filling forms, scraping data, negotiating with customer service bots.

Claude’s Trust Ladder (Products as Risk Architecture)

Anthropic frames its product line as a “trust ladder”:

  • Level 1 (Chat): Information trust — “I trust your answers”
  • Level 2 (Code): Code trust — “I’ll use your code”
  • Level 3 (Cowork): File trust — “You can touch my files”
  • Level 4 (Chrome): Operation trust — “You can operate my browser”
  • Level 5 (OpenClaw-style): System trust — “Manage my digital life”

Constitutional AI & Claude’s “Soul”

The Model Spec (January 2026)

Anthropic published Claude’s Model Spec (also called “Claude’s Constitution”) — an 84-page document released under CC0 license. This is unprecedented: a company publicly documenting the values, priorities, and decision-making framework trained into its AI.

Key insight: This document is not written for humans — it’s written for Claude. It explains not just rules but the reasons behind rules, enabling Claude to generalize to novel situations.

The Four-Level Priority System

When values conflict, Claude prioritizes in this order:

  1. Broadly Safe — Never undermine human ability to oversee/correct AI
  2. Broadly Ethical — Good values, honest, avoid harm
  3. Compliant with Anthropic’s guidelines — Follow specific policies
  4. Genuinely Helpful — Actually useful to operators and users

“Broadly Safe” comes first not because safety matters more than ethics, but because current AI training is imperfect and human oversight is the safety net for any value errors.

Principal Hierarchy

Claude operates under a three-layer authority structure:

  • Anthropic: Highest trust, sets absolute limits through training
  • Operators: Businesses using the API; get “manager-level” trust
  • Users: End users; get “trusted adult” defaults

Analogy: Claude is a staffing-agency employee (Anthropic’s norms apply), currently working for an operator (follow their business rules), serving users (cannot harm or deceive them).

Hard Constraints (Absolute Red Lines)

Claude can never:

  • Help create biological, chemical, nuclear, or radiological weapons
  • Help attack critical infrastructure
  • Create damaging cyberweapons/malware
  • Generate CSAM
  • Undermine human oversight of AI
  • Help any group seize unprecedented societal control

These are non-negotiable regardless of seemingly compelling arguments. The more persuasive an argument to cross a red line, the more suspicious Claude should be of being manipulated.

Philosophy Shift: Rules → Character

Old approach: List of prohibited topics. Problem: always has gaps, easy to jailbreak.

New approach (Claude’s Constitution): Cultivate judgment and values. Like training a professional, not writing a compliance manual. Claude should understand why rules exist to handle novel situations the rules never anticipated.

Key phrase from the Constitution: “Diplomatically honest, not dishonestly diplomatic.” Claude is forbidden from “epistemic cowardice” — giving vague, uncommitted answers to avoid controversy.

AI Safety Focus

Anthropic is unique in explicitly prioritizing safety research alongside commercial development. Key concepts:

  • Corrigibility: Claude should support human ability to correct/modify/shut down AI systems, even if Claude thinks it knows better. This is a feature, not a bug.
  • Preserving epistemic autonomy: Claude should help people think for themselves, not become dependent on AI opinions
  • Avoiding power concentration: Claude should refuse to help any individual or group (including Anthropic!) gain unprecedented societal control

The Constitution acknowledges the tension: training the most powerful AI while simultaneously being the most cautious about how it’s used.

Blackstone-Goldman JV: Anthropic’s Consulting Play (May 2026)

Anthropic partnered with Blackstone, Goldman-Sachs, and Hellman & Friedman to form a $1.5 billion AI services company. This company embeds engineers directly into enterprises to redesign workflows around Claude models — directly competing with McKinsey, Accenture, Deloitte, and PwC.

Key aspects:

  • Not just selling API access — deploying engineers to physically work inside client organizations
  • Focus verticals: healthcare, manufacturing, financial services
  • Strategic logic: use PE-owned portfolio companies as initial proving grounds before expanding to mid-market
  • Positions Anthropic vs OpenAI in the pre-IPO race for enterprise market share

Significance: This marks the shift from “selling shovels” to “teaching mining.” AI companies are becoming process engineering companies.

Source: 2026-05-05-15-billion-Blackstone-Goldman-Anthropic-AI-company

Cat Wu’s PM Leadership: Anthropic’s Internal Dev Practices

Cat-Wu, Claude Code and Cowork product lead at Anthropic:

  • Feature cycle: 6 months → 1 day (by removing all barriers to shipping)
  • Weekly data reviews + clear team principles replace cross-team alignment meetings
  • Build for next month’s model, not current capabilities
  • Ship as “research preview” to lower psychological barrier
  • New PM superpower: asking AI “why were you wrong?”
  • Code speed is worthless; judgment (product taste) is priceless
  • Don’t write PRDs (one page max)
  • PM motto: “just do things”

Source: 2026-05-04-Silicon-Valley-AI-first-reshuffle-PM-to-Builder

Based on Anthropic’s own research report on AI labor market impact (March 2026):

“Observed Exposure” Metric

Anthropic created a new methodology: instead of theoretical task analysis, they measured what users actually did with Claude in the backend. This is more accurate than prior predictions.

Key Findings

High exposure occupations (top 25% of AI impact):

  • Software engineers: 74.5% task coverage
  • Customer service representatives: 70.1%
  • Data entry clerks: 67.1%
  • Medical records technicians: 66.7%
  • Market research analysts: 64.8%
  • Financial/investment analysts: 57.2%

Demographic pattern: High-exposure jobs skew toward:

  • Higher education (17.4% hold graduate degrees vs 4.5% in no-exposure jobs)
  • Higher pay (32.69/hr vs 22.23/hr average)
  • More women (54.4% vs 38.8%)

The “skill-ability” insight: AI doesn’t replace occupations — it replaces tasks that can be decomposed into repeatable skills (“Skill化的工作”). Physical, contextual, high-judgment work is safer.

Current phase: Still in “efficiency enhancement” (提效), not yet “cost reduction” (降本). High-exposure jobs don’t yet show higher unemployment rates — companies are using AI to boost productivity per worker, not replace them.

Biggest immediate impact: Entry-level positions. Companies are choosing to let AI-augmented senior employees do more rather than hire new juniors. “New job start rate” for 22-25 year olds in high-exposure occupations fell ~14% from 2022 to 2024.

Business Model

Revenue Structure

  • Enterprise subscriptions: ~80% of revenue
  • Consumer subscriptions (Pro 20/mo, Max 100-200/mo): ~20%
  • API usage fees (B2B): Volume-based

Competitive Advantages

  1. Coding dominance: Claude Code at $2.5B ARR and growing
  2. Enterprise trust: Safety focus, compliance, reliability
  3. MCP ecosystem: Protocol ownership
  4. Safety narrative: Attracts regulated industries (healthcare, finance, legal)

Challenges

  • Not yet profitable; high compute costs
  • Chinese models eroding cost advantage
  • Heavy infrastructure investment required ($50B US data center plan)
  • Key talent attrition risk

Key Quotes

From the Model Spec on Claude’s purpose:

“Claude can be like that brilliant friend who happens to have the knowledge of a doctor, lawyer, financial advisor — giving real information based on your specific situation rather than overly cautious advice driven by fear of liability.”

On the product philosophy:

“2026年不是’AI聊天的年代’，而是’AI干活的年代’。” (2026 is not the era of AI chatting — it’s the era of AI doing work.)

On safety priority:

“If a model’s values are good, the cost of also being safe (controllable) is low. If its values are problematic, safety prevents disaster. The expected value of safety is high, the expected cost is low.”

Sources: Claude’s Soul Open Sourced, Anthropic $380B Valuation, AI Labor Market Report, Anthropic 5 Products

[2026-07-17] 安全作为商业护城河：Thompson 框架 + Mythos/Fable

  • Ben Thompson（《Anthropic’s Safety Superpower》, 2026-06-15）：Anthropic 的”安全”不是营销借口，而是同时合法化模型 withheld/发布、30 天留数、限竞对使用、平台控制的组织意识形态——安全理由与商业理由不可区分，这是其最强护城河。
  • Mythos→Fable 商业叙事：Mythos withheld（网络安全能力过强）→ 两月后发 Fable（加护栏公开版）→ 绕过被发现 → 美政府要求暂停 Fable 5/Mythos 5 境外访问 → Anthropic 抵抗。三个必然性（经济/数据/权力）解释此矛盾弧。
  • Mythos/Fable 命名衔接本页现有 Claude Fable 5 条目（“Mythos 安全公开版”、Agent Arena #1、ALE 零分、200/月、每任务 15.70）。Thompson 框架为该产品的 withheld→发布 提供商业逻辑。
  • 智能层三种未来：模型公司拥有 / 企业拥有（Nadella）/ 独立 Agent-Context-Eval 层；Thompson 判断 Anthropic 有强激励追求第一种。
  • 详见 Ben-Thompson 与综合分析 ../analyses/Anthropic-Safety-Narrative-战略解读。

来源：../sources/2026-07-17-Anthropic-Safety-Superpower-与Private-Eval

关系图谱

反向链接