Cognition

SWE-2模型发布，性能与前沿模型相当，成本降低70%。

软件正以计算速度重塑世界，Devin等编码代理使工程师效率提升10倍，推动行业变革。

Fireworks AI发布SWE-2模型，在保持前沿性能的同时将成本降低70%。

Devin Desktop和Devin CLI已集成GPT-6 Astra模型，其成本比Fable 5降低64%且性能接近。

Cognition宣布Devin的Fable 5.1版本通过缓存优化降低54%成本，Fusion系统执行成本降低47%，并为企业用户提供限时优惠。

Cognition宣布GPT-6 Astra将集成到Devin，其在FrontierCode 1.1上以64%更低的成本接近Fable 5性能。

AI Agent扩张将重塑算力需求与企业形态，终端用户付费意愿、网络安全风险升级和复合型公司崛起是关键变量。

GPT-5.6系列在FrontierCode平台实现价格性能突破，成本下降已自动同步至Devin用户。

模型路由在多模型部署中至关重要，NVIDIA、Cognition等公司正推动更智能的模型选择策略。

Kimi K3在Devin平台实现接近前沿性能，基准测试得分58.2%且环境管理能力突出。

Kimi K3在FrontierCode 1.1基准测试中展现接近前沿水平的工程能力，63.6%通过率与58.2%得分验证其实际应用潜力。

模型安全风险可通过后训练显著缓解，开源模型风险并非固有属性，行业协作推动AI安全工具创新。

构建生产级AI代理需聚焦上下文控制、确定性流程、状态管理和范围控制，基于Twelve-Factor Agents原则并结合多家公司实践。

Devin Desktop和CLI现已支持Inkling和Grok 4.5模型，并推出FrontierCode排行榜跟踪代码生成质量。

Cognition推出的FrontierCode排行榜现已上线，实时追踪能生成实际可合并代码的模型表现，包含Grok 4.5和Inkling等模型的评分及方法论。

Devin 现在在每次代码审查中自动进行安全审查，能发现扫描工具遗漏的漏洞并提供修复建议。

Cognition 公司推出 AI 生产力保证计划，投入 1000 万美元证明 Devin 能够提升工程产出。

One year after acquiring Windsurf, Cognition has fully integrated it into Devin Desktop, creating a unified platform for agent management and development.

异步代理时代正在重新定义软件开发，强调代理编排和端到端开发，而非单纯依赖开发者本地工作流。

文章探讨了异步代理技术如何推动工程实践变革，包括80%的AI提交和7倍PR增长的具体案例。

Vlad Feinberg's guide highlights mastering LLM kernel-level tuning and MoE architecture optimization as critical for entering frontier labs, while agent automation and observability emerge as infrastructure trends.

The optimal collaboration method between humans and AI agents has yet to be invented, as current session-based and prompt-driven interaction models have fundamental flaws. The Paperboy team is exploring more natural AI collaboration experiences.

OpenAI has announced the deprecation of its fine-tuning APIs, marking the end of the fine-tuning era. Despite this, top-tier AI companies like Cursor and Cognition are increasing their use of open model fine-tuning. Research benchmarks are becoming increasingly challenging, and agentic systems are making breakthroughs in science and mathematics.

Cognition closed a $1B Series D at a $26B valuation, becoming the largest remaining independent AI agent lab; ARR projected >$1B by EOY 2026; inference optimization is shifting to architectural improvements—EAGLE 3.1, vLLM, and Qwen3.5 significantly enhance long-context stability and throughput.

Sonnet 5模型在Devin Desktop/CLI中将临时减少30%配额消耗，8月31日后恢复原配额标准。

Cognition launches the AI Productivity Guarantee, pledging up to $10M to fund Devin usage if engineering value falls short of cost, shifting industry focus from tokens to tangible output.

Cognition launched an AI Productivity Guarantee that measures value by productive engineering hours saved per Devin task instead of token usage, offering up to $10M in usage funding if delivered value falls short of cost, shifting the industry from activity to output metrics.

文章探讨了在大规模代理开发中验证端到端测试能力的实践经验，强调了虚拟机技术的重要性。

Harvey's engineering team integrated their internal background agent Spectre into Devin Desktop, enabling organizational context to reside on engineers' laptops and flow across AI agents.

Cognition 宣布推出 Devin for Security，将 AI 工程师能力引入安全修复场景，应对 AI 加速漏洞利用与防御工具滞后之间的矛盾。