Anthropic Fable 5 携防护回归，GLM-5.2 编码能力比肩，Fable 首次单次生成 Megakernel
英文摘要
Anthropic restored access to Fable 5 with new cybersecurity guardrails, raised API rate limits, and expanded Claude Code artifacts to Pro/Max plans; Fable is expected to return to subscriptions when capacity permits. Open model GLM-5.2 reportedly reaches ~80% of Anthropic Sonnet 5's software-engineering capability at ~20% of the cost and is now usable within Claude Code via Hugging Face Inference Providers. In a landmark systems result, Elliot Arledge used Fable 5 to generate a single-launch megakernel for a Kimi-Linear decode workload, achieving an 18.7x speedup over the reference implementation and beating prior multi-kernel entries. The SWE-rebench leaderboard was updated with Claude Opus 4.8 xhigh at 56.5% solve rate, GLM-5.2 at 51.1%, and smaller open models like Qwen3.6-27B at 36.5%. The coding agent infrastructure is thickening with full-stack evals (Code Arena Fullstack) and agent-native parsing patterns; coordination, memory, and observability are now the bottlenecks.
中文摘要
Anthropic恢复Fable 5访问，新增网络安全分类器，提高API速率限制并向Pro/Max计划开放Claude Code artifacts；Fable将在容量允许后回归订阅。开源模型GLM-5.2在软件工程能力上达到Sonnet 5约80%，成本约20%，可通过Hugging Face Inference Providers在Claude Code中使用。Elliot Arledge用Fable 5生成首个单次完成的megakernel，针对Kimi-Linear解码负载实现18.7倍加速，超越此前多内核方案。SWE-rebench排行榜更新：Claude Opus 4.8 xhigh解决率56.5%，GLM-5.2 51.1%，小型开源模型如Qwen3.6-27B达36.5%。编码代理基础设施日趋完善，全栈评估（Code Arena Fullstack）和代理原生解析模式出现；协调、记忆和可观测性成为当前瓶颈。
关键要点
Anthropic redeployed Fable 5 with new cybersecurity classifiers, raised API rate limits, and expanded Claude Code artifacts; Fable returns to subscriptions once capacity allows.
Anthropic重新部署Fable 5，增加网络安全分类器，提高API速率限制并扩展Claude Code artifacts；Fable将在容量允许后回归订阅。
GLM-5.2 achieves roughly 80% of Sonnet 5 software-engineering capability at 20% of the price and is now usable in Claude Code via Hugging Face.
GLM-5.2软件工程能力达到Sonnet 5约80%，成本仅20%，可通过Hugging Face在Claude Code中使用。
Elliot Arledge used Fable 5 to write a single-launch megakernel for Kimi-Linear decode, hitting 18.7× speedup over baseline, demonstrating agentic kernel optimization.
Elliot Arledge用Fable 5生成单次megakernel，针对Kimi-Linear解码实现18.7倍加速，展示代理内核优化能力。
SWE-rebench updated: Claude Opus 4.8 xhigh at 56.5% solve rate, GLM-5.2 at 51.1%, and multiple smaller open models now listed.
SWE-rebench更新：Claude Opus 4.8 xhigh解决率56.5%，GLM-5.2 51.1%，并收录多个小型开源模型。
Full-stack coding evals gain traction: Code Arena’s Fullstack Code Arena extends evaluation to databases, API keys, and deployments, moving beyond frontend demos.
全栈编码评估兴起：Code Arena的Fullstack Code Arena将评估扩展至数据库、API密钥和部署，超越前端演示。
Memory management treated as a trainable skill in Stanford AutoMem paper yields 2–4× gains on Crafter, MiniHack, and NetHack.
Stanford AutoMem论文将记忆管理视为可训练技能，在Crafter、MiniHack和NetHack上实现2至4倍增益。