Anthropic 喺 2026 年 5 月嘅開發者大會上出奇地冇推新模型，反而全力晒冷平台級嘅代理功能，意味住佢嘅競爭策略，已經由鬥純粹嘅模型跑分，轉向打造識得自我管理、長期自主作業嘅 AI 系統 [36][39]。
What are the key details about Anthropic's Claude Code, including its origins as a research preview in February 2025, its evolution throughClaude Code evolved from a limited research preview into a billion-dollar agentic coding platform.
AI 提示
Create a landscape editorial hero image for this Studio Global article: What are the key details about Anthropic's Claude Code, including its origins as a research preview in February 2025, its evolution through. Article summary: Here is a comprehensive summary of what the search results confirm and where evidence is insufficient.. Topic tags: general, documentation, general web, user generated. Reference image context from search candidates: Reference image 1: visual subject "#### Learn On The Go. Build your Skills For a Better Tomorrow. Why Hyderabad is the Best Place to Learn Data Science in 2025. How Long Does It Take to Become a Full Stack Developer" source context "Claude AI Model Upgrades Explained: 2026 Overview" Reference image 2: visual subject "The image shows a timeline of Anthropic's news announcements and product releases from January to March 2026, highlighting key events
openai.com
Claude Code 最初係喺 2025 年 2 月 24 日，以一個有限度嘅研究預覽形式面世，當時同佢一齊登場嘅仲有 Claude 3.7 Sonnet 模型 。呢個係 Anthropic 第一個代理式編碼（Agentic Coding）產品，由 Day 1 開始就係設計畀開發者，等佢哋可以直接喺 Terminal 入面，將啲大規模嘅工程任務派畀 Claude 處理。
Anthropic 喺 2026 年 5 月嘅開發者大會上出奇地冇推新模型，反而全力晒冷平台級嘅代理功能，意味住佢嘅競爭策略，已經由鬥純粹嘅模型跑分，轉向打造識得自我管理、長期自主作業嘅 AI 系統 [36][39]。
What are the key details about Anthropic's Claude Code, including its origins as a research preview in February 2025, its evolution throughClaude Code evolved from a limited research preview into a billion-dollar agentic coding platform.
AI 提示
Create a landscape editorial hero image for this Studio Global article: What are the key details about Anthropic's Claude Code, including its origins as a research preview in February 2025, its evolution through. Article summary: Here is a comprehensive summary of what the search results confirm and where evidence is insufficient.. Topic tags: general, documentation, general web, user generated. Reference image context from search candidates: Reference image 1: visual subject "#### Learn On The Go. Build your Skills For a Better Tomorrow. Why Hyderabad is the Best Place to Learn Data Science in 2025. How Long Does It Take to Become a Full Stack Developer" source context "Claude AI Model Upgrades Explained: 2026 Overview" Reference image 2: visual subject "The image shows a timeline of Anthropic's news announcements and product releases from January to March 2026, highlighting key events
openai.com
Claude Code 最初係喺 2025 年 2 月 24 日，以一個有限度嘅研究預覽形式面世，當時同佢一齊登場嘅仲有 Claude 3.7 Sonnet 模型 。呢個係 Anthropic 第一個代理式編碼（Agentic Coding）產品，由 Day 1 開始就係設計畀開發者，等佢哋可以直接喺 Terminal 入面，將啲大規模嘅工程任務派畀 Claude 處理。
87.6% 呢個分數唔係固定不變嘅。佢取決於背後嘅模型、指令（Prompt）同埋「測試環境」（Harness）——即係嗰個負責編排點樣使用工具嘅執行環境。Claude Opus 4.7 嘅自適應模式會按每個任務嘅需要，動態咁分配運算資源，遇到複雜嘅重構就畀多啲資源。如果係冇開到呢個自適應模式嘅獨立版 Claude Code，喺同一個基準測試度就只會攞到 80.8% 。
喺更加難嘅 SWE-bench Pro 基準測試入面——呢個測試模擬嘅係更難嘅真實世界問題——Opus 4.7 攞到 64.3%，領先 GPT-5.4 嘅 57.7%、GPT-5.5 嘅 58.6% 同埋 Gemini 3.1 Pro 嘅 54.2% 。後嚟 Opus 4.8 再將 SWE-bench Pro 嘅分數推到上 69.2% 。
其他基準測試同競爭對手嘅表現
Claude Code 喺多個基準測試度嘅表現都好標青：
Terminal-Bench 2.0：用 Opus 4.7 攞到 69.4% 。
CursorBench：用 Opus 4.7 攞到 70%，對比 Opus 4.6 嗰陣係 58% 。
87.6% 呢個分數唔係固定不變嘅。佢取決於背後嘅模型、指令（Prompt）同埋「測試環境」（Harness）——即係嗰個負責編排點樣使用工具嘅執行環境。Claude Opus 4.7 嘅自適應模式會按每個任務嘅需要，動態咁分配運算資源，遇到複雜嘅重構就畀多啲資源。如果係冇開到呢個自適應模式嘅獨立版 Claude Code，喺同一個基準測試度就只會攞到 80.8% 。
喺更加難嘅 SWE-bench Pro 基準測試入面——呢個測試模擬嘅係更難嘅真實世界問題——Opus 4.7 攞到 64.3%，領先 GPT-5.4 嘅 57.7%、GPT-5.5 嘅 58.6% 同埋 Gemini 3.1 Pro 嘅 54.2% 。後嚟 Opus 4.8 再將 SWE-bench Pro 嘅分數推到上 69.2% 。
其他基準測試同競爭對手嘅表現
Claude Code 喺多個基準測試度嘅表現都好標青：
Terminal-Bench 2.0：用 Opus 4.7 攞到 69.4% 。
CursorBench：用 Opus 4.7 攞到 70%，對比 Opus 4.6 嗰陣係 58% 。
