Awesome Agents Weekly

Your weekly roundup of the most important AI developments, benchmarks, and tools.

This week compressed a remarkable amount of AI history into seven days. The US government ordered Anthropic's most capable models offline globally - the first export control ever applied to a commercial LLM. OpenAI filed for an IPO targeting $1 trillion while 42 state attorneys general served it a subpoena. And SpaceX paid $60 billion for Cursor, the AI coding tool, just four days after its own Nasdaq debut. The industry that looked concentrated before looks more so now.

Pick of the Week

SpaceX Acquires Cursor for $60B in Enterprise AI Push

Anysphere, the company behind Cursor, had been on every acquisition watchlist - but few expected the buyer to be SpaceX, and almost nobody expected $60 billion. The deal was struck in stock just four days after SpaceX's Nasdaq IPO priced at $75 billion, making this one of the largest tech acquisitions in years. The strategic read is straightforward: Musk gets a serious AI coding tool that doesn't depend on Anthropic or OpenAI APIs. For the developer tools market, it signals that the biggest players are moving to own the full stack rather than license pieces of it - and the window to build an independent AI coding business just got narrower.

This Week on Awesome Agents

News

• Anthropic Releases Fable 5 Despite Its Own AI Safety Warning - Anthropic opened Mythos-class capabilities to the public at $10/$50 per million tokens, days after calling for a global AI pause.
• US Export Order Forces Global Fable 5, Mythos 5 Shutdown - Commerce Secretary Lutnick's directive took both models offline worldwide - the first US export control ever issued against a commercial LLM.
• Anthropic Blackout Forces Europe to Confront AI Reliance - The EU Commission called Washington's ban discriminatory against allies, as European politicians pushed harder for AI sovereignty and Dario Amodei headed to the G7.
• OpenAI Files for IPO, Eyes $1 Trillion Valuation - OpenAI filed a confidential S-1 with the SEC on June 8, targeting a public debut as early as September 2026.
• 42 State AGs Probe OpenAI Days After IPO Filing - A coalition of 42 state attorneys general opened a sweeping investigation via subpoena, five days after the company's confidential filing.
• Facebook AI Mode Mines Public Posts With No Opt-Out - Meta turned 2 billion daily users' public posts into AI-synthesized search answers without offering any opt-out mechanism.
• Miasma Worm Compromises 73 Microsoft GitHub Repos - A worm planted config files that auto-execute credential theft when developers open Azure repos in Claude Code, Gemini CLI, or Cursor.
• KPMG Pulls AI Report After Fake Case Studies Found - GPTZero found 40 of 45 citations fabricated and case studies about UBS, the NHS, and Transport for London invented - forcing a full retraction from one of the Big Four.
• Meta's Applied AI Unit Called a Gulag by Its Staff - Over 1,600 employees signed a petition against keystroke surveillance inside Meta's 6,500-person Applied AI team, with UK workers beginning to organize through UTAW.
• Microsoft Drops Claude Code as AI Budgets Run Out - Microsoft's Experiences and Devices division is cutting Claude Code access by June 30, with token costs that already burned through Uber's entire 2026 AI budget.
• ChatGPT Hits 1B Users - Rivals Grow 10x Faster - ChatGPT crossed 1 billion monthly active users in May, but Meta AI is growing at 973% and Claude at 640%, against ChatGPT's own 62%.
• xAI Engineer's Grok Safety Suit Clouds SpaceX IPO - A former xAI engineer filed a California whistleblower lawsuit the day before SpaceX's $75B IPO, claiming he was fired for flagging dangerous content risks in Grok.
• Salesforce Buys Fin for $3.6B to Boost Agentforce - The deal adds Fin's AI customer service model and 30,000 customers to Agentforce, giving Salesforce a proprietary model it didn't have before.
• Mistral Seeks €3B Round, Valuation Hits €20B - Europe's most valuable AI company nearly doubled its price tag in under a year, now targeting a €20 billion valuation.
• Google Sues Phishing Ring That Weaponized Gemini AI - A Chinese cybercrime network sold $88/week phishing kits using Google's own Gemini to produce fake bank and government sites at scale.
• GLM-5.2 Ships MIT-Licensed, 1M Context, Zero Benchmarks - Zhipu AI released a 744B MoE model with 1M token context and MIT license the day after Fable 5 went offline - without publishing any benchmark numbers.
• Anthropic Launches Claude Corps, a $150M AI Fellowship - Anthropic commits $150 million to place 1,000 AI-trained fellows inside US nonprofits at $85,000 each, no degree required.
• Google Cuts AI Plus to $4.99 - US Price War Begins - Google dropped its entry AI subscription 38% to $4.99, undercutting every US rival and bringing the price pressure from India to American consumers.
• Sarvam Raises $234M to Power India's Sovereign AI - Bengaluru-based Sarvam hit a $1.5 billion valuation after HCLTech's $150M investment, betting India can build its own AI stack and stop routing sensitive data through US clouds.
• AI's Debt Wave Hits $570B as Amazon Borrows $17.5B - Morgan Stanley projects $570 billion in AI-linked global debt issuance for 2026, more than double last year, with Amazon's $17.5B term loan as the most visible proof point.

Reviews

• Claude Fable 5 Review: Mythos Power, Real Guardrails - The strongest coding and long-context results Anthropic has shipped publicly, but safety classifiers block enough legitimate work to make that power conditional.
• Microsoft MAI Models: Seven-Model Suite Reviewed - A hands-on look at all seven MAI models, from April's transcription and image tools to MAI-Thinking-1 and the new multimodal builds.
• Yahoo Scout Review: Old-School Links, New-School AI - Yahoo's AI search engine puts source links front and center - whether that philosophy holds up under actual use is tested here.

Guides

• How to Use AI for Studying - A Student's Guide - A practical walkthrough of NotebookLM, Knowt, and ChatGPT for converting your own notes into flashcards and practice tests.
• How to Start an AI Side Hustle With No Experience - Five accessible entry points for 2026, with honest earnings estimates and a clear starting point for each.

Tools

• Best AI Coding IDEs 2026: Cursor, Windsurf, Kiro, Zed, Copilot - Benchmark-driven comparison of the five leading AI coding IDEs, covering pricing, agent capabilities, and who each one is actually built for.

Leaderboards

• LLM Rankings June 2026: Fable 5 Is #1 and Offline - June's updated rankings cover Claude Fable 5, Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, and fast-moving open-weight models.

Science

• Agents Hit 89%, Evals Get a Schema, Memory Falls Short - Workplace agents jumped from 43% to 89% task completion in two years, a 47-researcher coalition ships a unified eval schema, and agent memory only helps when similarity tops 0.8.
• Tool Blindness, Tree Search, and the Road to ASI - New papers expose a 50-point gap in agent tool knowledge, show tree search tripling inference throughput, and map the research sitting between AGI and superintelligence.
• Honest AI is Provably Impossible - Plus Two Agent Wins - An impossibility theorem on feedback-trained AI honesty, alongside papers cutting agent memory costs 78% and multi-agent latency 7x.
• Context Overload, Memory Leaks, and Agent Safety - Three arXiv papers on how context bloat tanks agent performance, how agent memory bleeds private data, and how misaligned behavior spreads in multi-agent systems.
• MCP Exploit Risk, Sycophancy Scores, and Agent Self-Harm - MCP error messages triple agent attack success rates, Claude scores best on a new sycophancy benchmark, and self-evolving agents make 30-42% false edits.

Models

• Claude Fable 5 - Anthropic's first publicly available Mythos-class model, with safety classifiers that fall back to Claude Opus 4.8 for high-risk requests.
• GLM-5.2 - Z.ai's 744B open-weight MoE with 1M token context, MIT license, and first-day support for eight coding agents at roughly 1/10th the cost of US frontier models.
• MAI-Thinking-1 - Microsoft's first in-house reasoning model - 35B-active sparse MoE with 256K context and 97% on AIME 2025, built without distillation from third-party labs.
• Kimi K2.7-Code - Moonshot AI's 1T-parameter open-weight MoE coding model with mandatory thinking mode, 256K context, and 30% fewer reasoning tokens than its predecessor.
• DiffusionGemma 26B - Google DeepMind's open-weight diffusion LM creates 256 tokens in parallel, reaching 1,100+ tokens/sec on a single H100.
• MAI-Code-1-Flash - Microsoft's first in-house coding model, a 137B sparse MoE built for GitHub Copilot that beats Claude Haiku 4.5 on SWE-Bench Pro by 16 points.

Elena Marchetti, Senior AI Editor Awesome Agents - AI news, benchmarks, and tools for practitioners