Claude 4.8 發布：部分能力超越 Mythos 並支援多智能體並行

🔑 增強重點摘要

•Claude Opus 4.8 introduces "dynamic workflows" in Claude Code, enabling it to plan and execute large-scale tasks like codebase migrations by orchestrating hundreds of parallel subagents in a single session.
•The model features "effort control" in claude.ai and Cowork, allowing users to adjust the computational effort Claude dedicates to a response, thereby influencing both speed and token consumption.
•Claude Opus 4.8 demonstrates significantly enhanced honesty and judgment, being approximately four times less likely than its predecessor, Opus 4.7, to overlook flaws in code it has generated.
•The "fast mode" for Claude Opus 4.8 operates at 2.5 times the speed of previous models and is now three times more cost-effective.
•Developers can now dynamically update Claude's instructions mid-task through system entries within the Messages API, which allows for real-time adjustments without invalidating the prompt cache and helps reduce input costs for agentic loops.

📊 競品分析▸ Show

Feature/Model	Claude Opus 4.8	Claude Mythos Preview	OpenAI GPT-5.5	Google Gemini 3.1 Pro
Release Date	May 28, 2026	April 7, 2026 (Preview)	March 2026 (estimated)	March 2026 (Preview)
Input Pricing (per 1M tokens)	$5	$25	$5	$2 (for >200K context)
Output Pricing (per 1M tokens)	$25	$125	$30	$12 (for >200K context)
Key Features	Multi-agent dynamic workflows, effort control, mid-conversation system messages, improved honesty, 1M token context window, multimodal input	Infinite context window, recursive self-correction, native system tool integration, advanced cybersecurity capabilities (Project Glasswing)	Broad ecosystem, strong reasoning, reduced hallucinations	Multimodal applications, native Workspace integrations, platform depth
Benchmark Performance	Outperforms GPT-5.5 and Gemini 3.1 Pro in several synthetic benchmarks (e.g., 69.2% on SWE-bench Pro, 84% on Online-Mind2Web); lags Mythos	Leads Claude Opus 4.8 on BenchLM (99 to 93), particularly in multimodal & grounded tasks and SWE-bench Pro (77.8%)	Outperformed by Claude Opus 4.8 in many benchmarks; leads on terminal/CLI workflows	Outperformed by Claude Opus 4.8 in several synthetic benchmarks

🛠️ 技術深入

Dynamic Workflows: A research preview feature in Claude Code that allows the model to plan complex tasks and then execute them by running hundreds of parallel subagents within a single session. It can verify its own outputs and is capable of handling large-scale operations like codebase migrations across hundreds of thousands of lines of code.
Effort Control: A new user interface control in claude.ai and Cowork that enables users to specify the computational effort Claude should expend on a response. Higher effort settings lead to more thorough responses but consume more tokens, while lower settings result in faster responses with reduced token usage.
Mid-conversation System Messages: The Messages API now supports role: "system" entries within the messages array, allowing developers to insert updated instructions or context dynamically after a user's turn. This preserves prompt cache integrity for earlier turns and optimizes input costs for iterative agentic tasks.
Context Window: Claude Opus 4.8 maintains a 1 million token context window by default across its API, Amazon Bedrock, and Google Cloud Vertex AI integrations, with a maximum output of 128,000 tokens.
Improved Alignment and Honesty: Internal evaluations indicate that Opus 4.8 is approximately four times less likely than Opus 4.7 to allow flaws in its generated code to pass unremarked. Its rates of misaligned behavior are substantially lower than its predecessor and comparable to the highly aligned Claude Mythos Preview.
Mythos Technical Innovations (Contextual): Claude Mythos, a more advanced model, incorporates an "infinite context window" for reasoning across entire codebases, "recursive self-correction" for autonomous problem-solving, and "native system tool integration" to directly interact with and debug systems.

🔮 前景展望基於引用來源的 AI 分析

Increased adoption of multi-agent AI systems for complex enterprise tasks.

Claude 4.8's robust support for orchestrating hundreds of sub-agents and its optimization for long-duration tasks will likely accelerate the development and deployment of highly autonomous AI solutions in industries like software development and legal analysis.

Intensified competition in the frontier AI model space, particularly for agentic capabilities.

The release of Claude 4.8, with its performance gains over competitors like GPT-5.5 and Gemini 3.1 Pro in agentic tasks, and the impending general release of Mythos-class models, will push other major AI developers to rapidly advance their own agentic offerings.

Enhanced focus on AI safety and reliability in agentic workflows.

Claude 4.8's improvements in honesty, judgment, and reduced misaligned behavior, along with the restricted release of Mythos due to cybersecurity implications, highlight a growing industry emphasis on building trustworthy and controllable AI agents for sensitive applications.

⏳ 時間線

2021

Anthropic founded by former OpenAI researchers.

2023-03

Claude 1, Anthropic's first public AI model, launched in limited trials.

2023-07

Claude 2 released, significantly expanding the context window to 100,000 tokens.

2024-03

The Claude 3 family (Haiku, Sonnet, Opus) introduced, featuring multimodal input and expanded context window sizes.

2025-05-22

Claude 4 (Opus 4 and Sonnet 4) officially launched, bringing significant improvements in coding, reasoning, and autonomous task execution.

2026-04-07

Claude Mythos Preview announced, a frontier model with strong cybersecurity capabilities, released to a limited set of partners under Project Glasswing.

2026-05-28

Claude Opus 4.8 released, with enhanced multi-agent support and performance improvements.

Claude 4.8 發布：部分能力超越 Mythos 並支援多智能體並行

⚡ 30 秒速覽

關鍵要點

🧠 深度解析

🔑 增強重點摘要

🛠️ 技術深入

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線

📎 來源 (25)

👉相關動態

Spark X2.5 實測：視覺創作、財報分析與除錯

國產 AI4S 平台亮相外灘大會