🗾ITmedia AI+ (日本)•較早收集於 10h
Cursor 發布 Composer 2.5，實現高效能且低成本編碼
💡全新編碼模型以十分之一成本達到頂尖效能，是提升開發效率與節省預算的重大突破。
⚡ 30-Second TL;DR
有什麼變化
Composer 2.5 以十分之一的成本實現頂尖模型效能
為什麼重要
此發布大幅降低了高品質 AI 輔助編碼的門檻，可能對現有編碼代理供應商的定價模式造成衝擊。
下一步行動
將 Composer 2.5 與您目前的編碼代理工作流程進行基準測試，評估是否能在不犧牲品質的前提下降低 API 成本。
誰應關注：Developers & AI Engineers
關鍵要點
- •Composer 2.5 以十分之一的成本實現頂尖模型效能
- •在 Artificial Analysis 編碼代理基準測試中位列第三
- •效能表現足以媲美 Claude Opus 4.7 與 GPT-5.5
🧠 深度解析
Web-grounded analysis with 16 cited sources.
🔑 增強重點摘要
- •Composer 2.5, released on May 18, 2026, is built upon Moonshot's open-source Kimi K2.5 checkpoint, with Cursor contributing approximately 85% of the total compute through its own additional training and reinforcement learning.
- •The model is offered in two variants: a standard version priced at $0.50 per million input tokens and $2.50 per million output tokens, and a 'Fast' variant, which is the default in Cursor, costing $3.00 per million input tokens and $15.00 per million output tokens.
- •Composer 2.5 demonstrates substantial performance improvements over its predecessor, Composer 2, achieving a 14-point increase on the Artificial Analysis Coding Agent Index (from 48 to 62) and notable gains on other benchmarks like SWE-Bench Pro (+35 points).
- •The training regimen for Composer 2.5 involved 25 times more synthetic tasks than Composer 2, incorporating targeted textual feedback for reinforcement learning to refine specific behaviors such as tool use, communication style, and effort calibration.
- •Despite its advanced capabilities, Composer 2.5 is positioned as significantly more cost-effective, with its standard variant costing as little as $0.07 per task and the Fast variant $0.44 per task, making it 10-60 times cheaper than higher-effort versions of Claude Opus 4.7 and GPT-5.5 on the Artificial Analysis Coding Agent Index.
📊 競品分析▸ Show
| Feature/Metric | Cursor Composer 2.5 | Claude Opus 4.7 | GPT-5.5 |
|---|---|---|---|
| Release Date | May 18, 2026 | April 16, 2026 | April 23, 2026 |
| Artificial Analysis Coding Agent Index | 62 | 66 (max variant in Claude Code) | 65 (xhigh reasoning in Codex) |
| SWE-Bench Multilingual / Verified | 79.8% (SWE-Bench Multilingual) | 87.6% (SWE-bench Verified) | 58.6% (SWE-Bench Pro) |
| CursorBench v3.1 | 63.2% | 70% | N/A |
| Terminal-Bench 2.0 | Improved (+2 points over Composer 2) | 69.4% | 82.7% |
| Pricing (Input/Output per 1M tokens) | Standard: $0.50 / $2.50 Fast: $3.00 / $15.00 | Standard: $5.00 / $25.00 Fast: $30.00 / $150.00 | Standard: $5.00 / $30.00 Pro: $30.00 / $180.00 |
| Cost per Task (Artificial Analysis) | Standard: $0.07 Fast: $0.44 | $4.10 (max variant in Claude Code) | $4.82 (xhigh reasoning in Codex) |
| Context Window | N/A | 1M tokens | 1M tokens, 1.1M tokens |
| Tokenizer Impact | N/A | New tokenizer may increase token counts by up to 35% for the same text | More token-efficient than GPT-5.4 |
| Availability | Exclusively in Cursor IDE and Cursor CLI (no external API) | Claude products, Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry | ChatGPT Plus, Pro, Business, Enterprise subscriptions; API access |
🛠️ 技術深入
- Composer 2.5 is built on Moonshot's open-source Kimi K2.5 checkpoint.
- Cursor reports that approximately 85% of the total compute for Composer 2.5 came from its own additional training and reinforcement learning.
- The model was trained on 25 times more synthetic tasks compared to its predecessor, Composer 2.
- A key aspect of its training involves targeted textual feedback for reinforcement learning, which helps to precisely shape specific behaviors such as tool use, communication style, and effort calibration.
- Composer 2.5 is specifically tuned for long-horizon tasks, which involve multi-step, context-heavy workflows.
🔮 前景展望AI analysis grounded in cited sources
Cursor will significantly expand its AI model capabilities through a strategic partnership with xAI.
Cursor is actively training a much larger successor model from scratch with SpaceX and xAI, leveraging ten times more compute on the Colossus-2 cluster, indicating a major leap in future capability.
The competitive landscape for AI coding agents will intensify, potentially leading to further price reductions or feature bundling by competitors.
Composer 2.5's ability to match top-tier performance at a fraction of the cost puts significant pressure on competitors like Anthropic and OpenAI, who may need to adjust their strategies to remain competitive.
⏳ 時間線
2022
Anysphere (Cursor's parent company) founded by MIT graduates Michael Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger.
2023
Cursor, an AI-first code editor, was launched.
2023-10
Raised an $8 million seed round led by the OpenAI Startup Fund.
2025-11-13
Closed a $2.3 billion Series D funding round, co-led by Accel and Coatue Management, valuing the company at $29.3 billion, with participation from Google and Nvidia.
2026-04-21
xAI announced an optional deal to acquire Anysphere (Cursor) for $60 billion or pay $10 billion for collaborative work.
2026-05-18
Launched 'Composer 2.5', its latest high-performance, low-cost AI coding model.
📎 來源 (16)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： ITmedia AI+ (日本) ↗