2026-06-13 · Sat
generated 10:34:03
🌟 Today's Headline
Perplexity Launches Search as Code for AI Agent Orchestration
Perplexity unveiled 'Search as Code,' a fundamental rearchitecture of how AI agents execute search tasks. Rather than a new model, this transforms search from a black-box API call into composable building blocks that agents orchestrate per task. On a benchmark of 230+ cybersecurity vulnerabilities, Search as Code achieved 100% accuracy versus under 25% for competing systems including GPT-5.5 with web search. The token cost dropped 85%, meaning equivalent results with far fewer API calls. This infrastructure-level shift quietly changes which applications are economically viable. Previously unaffordable use cases due to cost or latency constraints now become profitable. For creators: infrastructure and architectural improvements often reshape the playing field more than model parameter increases. When you see an AI system redesigned rather than just enlarged, understand the underlying logic.
💬 Editor's Note
The real innovation isn't model scale—it's search architecture. An 85% token reduction while beating GPT-5.5 on precision signals a fundamental shift in bottlenecks: raw reasoning is abundant, efficient information retrieval is now the moat.
🔥Today's Highlights
01
Claude Fable 5 Released: Anthropic's Most Capable Model with Advanced Vision and Coding
10/10
New Product
Anthropic released Claude Fable 5, its most capable public model. It achieves 80.3% on SWE-Bench Pro (vs GPT-5.5's 58.6%) and outperforms Opus 4.8 by over 10% on multiple tasks. Stripe demonstrated its power completing a 50-million-line Ruby codebase migration in one day—work that typically requires two months.
02
25 Open-Weight Models Released in Single Week Across All Modalities
10/10
Industry
This week witnessed the densest release window for open-weight models on record: 25 new models across all modalities—text chat, image generation, speech, music, video, and 3D—launched within a single week. This reflects the open-source AI community's vitality and marks a new phase of model democratization. Simultaneously, proprietary model pricing diverges sharply.
9/10
News
The US government has issued an export control directive suspending all access to Anthropic's Fable 5 and Mythos 5 models by foreign nationals, citing national security authorities. This restriction applies universally, including foreign national Anthropic employees, creating significant operational and international implications for AI development.
9/10
New Product
Google Cloud introduces the Open Knowledge Format to provide foundation models with relevant context for accurate, actionable outputs. As models advance, contextual information becomes critical for agentic system performance. The format enables better information integration for code writing, document analysis, and data summarization tasks.
9/10
New Product
Microsoft announced Azure Container Apps Sandboxes in public preview, enabling developers to safely execute untrusted code generated by AI agents in hardware-isolated environments. Each sandbox boots in under one second from OCI disk images and auto-scales to thousands of instances while costing nothing when idle.
8/10
News
French AI startup Mistral AI negotiates a new funding round of approximately 3 billion euros at a valuation around 20 billion euros. This major funding round demonstrates Europe's strategic commitment to building competitive AI capabilities independent of US-dominated tech companies.
📊Topic Clusters
📌 Anthropic Fable 5
新模型发布后面临透明度危机和美国出口管制禁令，引发行业关注
Claude Fable 5 Released: Anthropic's Most Capable Model with Advanced Vision and Coding 10
Anthropic Apologizes for Secretly Restricting Claude Fable 5, Commits to Real-Time Transparency 8
📌 AI融资与IPO热潮
Mistral融资30亿欧元，Anthropic和OpenAI IPO在即，掀起资本竞逐
Bezos' Prometheus Raises $12B for 'Artificial General Engineer' 8
📌 Meta AI危机
6500人AI部门混乱，员工反对Zuckerberg Hackathon，组织陷入混沌
📌 开源模型爆发
一周25+开源模型发布，Kimi、olmo、Ollama等发起攻防，与闭源模型竞争
25 Open-Weight Models Released in Single Week Across All Modalities 10
Kimi.ai open-sources new coding model with 30% fewer reasoning tokens 8
📖Worth a Deep Read
🕐 ~10 min read
· Tech
8/10
Kimi.ai open-sources new coding model with 30% fewer reasoning tokens
💡 Detailed technical reference
Moonshot AI unveiled Kimi-K2.7-Code, a new open-source coding model designed to be faster, smarter, and more cost-effective than previous versions. The key innovation is a 30% reduction in reasoning tokens — a significant efficiency gain that directly translates to lower costs and reduced latency for developers. The model was specifically trained to avoid 'overthinking,' a problem where AI systems consume excessive tokens while reasoning through programming problems. Each token costs money and adds computational delay, so reducing token usage while maintaining or improving output quality is a major breakthrough. By cutting unnecessary reasoning, the model can deliver the same or better code suggestions faster and cheaper. This positions it as a competitive alternative to proprietary coding models like GPT-4 or Claude for open-source projects and developers who want to avoid vendor lock-in. The open-source release means developers can run it locally, fine-tune it for their specific use cases, and benefit from community contributions.
🕐 ~3 min read
· Opinion
7/10
💡 Views and arguments worth studying
Andrew Singleton uses satirical storytelling about crematorium and propane company investments to critique how AI industry returns are reported. The allegory illustrates how financial transactions can manipulate numbers to create illusions of massive returns, reflecting broader skepticism about inflated AI investment gains and valuation metrics.
🕐 ~3 min read
· Industry
7/10
💡 Industry trends and analysis
美国商务部上周五以国家安全为由，要求Anthropic暂停所有外国国民（含公司内部外籍员工）对Fable 5和Mythos 5的访问。Anthropic已紧急对所有客户禁用这两个模型。起因是有人发现一种jailbreak可诱导模型提供本应拒绝的网络安全帮助。Anthropic认为政府未展示通用jailbreak，该技术范围狭窄，仅发现少量已知小漏洞，且其他公开模型也能提供类似能力。商务部长Howard Lutnick称这些模型将面临出口限制，直至美国政府强化国家安全系统（预计未来几周内）。Anthropic表示完美抵抗jailbreak目前任何模型都难以实现，并称此为误解，正努力恢复访问。其他Claude模型不受影响。
🕐 ~3 min read
· Tutorial
7/10
💡 Can be adapted into tutorial material
在OpenRouter上追加`：floor`可获取最便宜提供商，通过`max_price`设定花费上限，并可免费使用20多个零成本模型。同时需注意避免计费陷阱。
🕐 ~3 min read
· Tutorial
6/10
💡 Can be adapted into tutorial material
Hugging Face releases olmo-eval, an evaluation workbench designed to streamline the model development loop. The tool provides a systematic framework for developers to evaluate models during development, improving iteration speed and model quality through structured evaluation processes.
📂Browse by Category
New Product
OpenAI launches three new Academy courses designed to help people develop practical AI skills, create repeatable workflows, and apply AI agents in everyday work. The courses address growing demand for structured learning in the emerging era of AI-integrated workplaces.
OpenAI introduces flexible rate-limit resets for Codex users on Go, Plus, Pro, and Business plans. Users can now bank resets and trigger them manually instead of fixed expiration schedules, enabling immediate usage continuation when hitting caps mid-session. This flexibility addresses pricing friction in competitive AI agent markets.
Pinecone announced integration between its Nexus knowledge engine and Microsoft OneLake, enabling enterprise AI agents to directly access and reason over corporate data. This integration transforms how enterprise organizations deploy agentic systems over proprietary information assets.
Opinion
Anthropic's Claude Fable 5 has been called 'the best coding model in the world' by Every, but Dan Shipper argues most knowledge workers are missing its real significance. For casual users, Fable 5 feels incremental — it costs twice as much but doesn't dramatically improve everyday tasks.
Latent Space highlights Loopcraft, a concept from researchers Peter Steinberger, Boris Cherny, and Andrej Karpathy exploring patterns for stacking loops in AI systems. The concept provides insights into structuring recursive and iterative processes in AI agent design, informing sophisticated control flow architectures.
Ed Zitron observes that OpenAI and Anthropic are approaching IPO announcements despite burning billions annually with no clear profitability path. Both companies' public offerings signal potential end of the current tech era and raise critical questions about sustainable AI business models and investor rationality.
Industry
Anthropic is throttling its Mythos model for certain tasks while building competing applications, triggering pushback from customers, partners, and investors. This mirrors Microsoft's historical challenge of balancing platform neutrality with proprietary product development, revealing fundamental tensions in platform-based AI businesses.
OpenAI 正被一个由多州总检察长组成的联盟调查，该联盟已向这家人工智能公司索取涵盖广泛主题的信息。
GOP lawmakers and tech investors have attributed anti-data-center sentiment in the US to Chinese interference, but experts argue the situation is more nuanced and complex. Local environmental concerns, power supply constraints, and community opposition play equally significant roles in the data center resistance movement.
Tech
OpenAI lets Codex users bank and save rate limit resets for later
8
OpenAI has introduced a new feature for Codex users: the ability to bank and schedule rate limit resets on demand. Previously, rate limit resets were automatic and often occurred at inconvenient times like 3am. Now users can trigger them whenever they need, typically during heavy coding sessions. Every Go, Plus, Pro, and Business user receives one free banked reset immediately.
Tutorial
Simon Willison enhanced his OpenAI WebRTC audio tool with document context support. Originally built in December 2024 to test OpenAI's realtime audio API, this update adds document context capability, enabling richer interactive conversations and expanding the tool's utility for document-based workflows.
Anthropic对近5.2万美国人调查显示：48%将治愈癌症等疾病列为首要期望，36%希望AI帮助残障人士。64%担忧AI导致失业，56%担忧认知依赖，52%担忧信息误导。超70%支持政府监管，最关注隐私（56%）、儿童安全（52%）和责任归属（49%）。仅15%信任AI公司决策。
AgentPerf, the industry's first agentic AI benchmark from Artificial Analysis, establishes standardized metrics for comparing agent-based AI systems. NVIDIA Blackwell Ultra NVL72 leads initial results, delivering 20x more agents per megawatt than previous NVIDIA generations, setting new performance standards for enterprise agentic AI infrastructure.
📭Skip Today
Auto-filtered. Here's why — so you know you're not missing out:
The 500-Dollar AI Prompt, Now Free: The 'Idea Miner' Five-Stage Business Launch System
→ Already covered, no new facts today
OpenAI WebRTC Audio Session, now with document context
→ Already covered, no new facts today
New OpenAI Academy courses for the next era of work
→ Already covered, no new facts today
Over half of Americans fear losing both their jobs and their independent thinking to AI, survey finds
→ Already covered, no new facts today
OpenAI kicks off the AI price wars with flexible rate-limit resets for its Codex coding agent
→ Already covered, no new facts today
The AI industry's platform trap is starting to look a lot like Microsoft's
→ Already covered, no new facts today
v2.1.177
→ Already covered, no new facts today
v2.1.176
→ Already covered, no new facts today
📎 Long Tail (55) · click to expand
The 500-Dollar AI Prompt, Now Free: The 'Idea Miner' Five-Stage Business Launch System 4