Awesome Agents Weekly

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

Three stories defined this week's coverage. The White House banned Anthropic's Fable 5 with a condition security researchers say can't exist: a jailbreak-proof language model. SpaceX bought AI coding tool Cursor for $60 billion, four days after its Nasdaq IPO. And for the first time since ChatGPT launched, OpenAI's market share dropped below 50%. The infrastructure bets ran alongside all of it - $6.3B in compute deals, $310M into world models, and a Canadian pension fund committing a billion dollars to Indian data centers.

Pick of the Week

Fable 5 Is Banned Over a Problem That Can't Be Solved

The White House won't lift its ban on Anthropic's Fable 5 until the model can be made jailbreak-proof. Security experts explain why that condition is technically impossible - not with more safety training, not with better filters, not with any approach available today. Asking for a jailbreak-proof language model is closer to asking for a waterproof sponge than setting a hard engineering target. The policy demand and the technical reality are pointed in opposite directions, and until someone bridges that gap, two of Anthropic's best models sit offline worldwide.

This Week on Awesome Agents

News

• Fable 5 Is Banned Over a Problem That Can't Be Solved - The White House ban on Fable 5 depends on a jailbreak-proof guarantee that security researchers say can't exist.
• How Amazon CEO Triggered the Fable 5 Shutdown - Andy Jassy flagged a Fable 5 jailbreak on a routine White House call, triggering a 90-minute ultimatum that took down Anthropic's two best models worldwide.
• Reflection AI Locks $6.3B Compute Deal With SpaceX - The AlphaGo co-founder's open-source lab will pay SpaceX $150M monthly for Nvidia GB300 chips at Colossus 2, before shipping a single public model.
• SpaceX Acquires Cursor for $60B in Enterprise AI Push - SpaceX signed a deal to acquire Anysphere, maker of AI coding tool Cursor, for $60B in stock just four days after its record-breaking Nasdaq IPO.
• ChatGPT Falls Below 50% Market Share for First Time - Sensor Tower's State of AI 2026 report shows ChatGPT's True Audience share fell to 46.4% in May as Gemini and Claude eat into OpenAI's lead.
• Anthropic Surveyed 52K Americans - Just 15% Trust AI - Anthropic's first Public Record survey of 51,993 Americans finds 64% fear job displacement, only 15% trust AI companies, and 70%+ support government regulation.
• AI CEOs at G7 Call for US-Led Global Standards Forum - Altman, Amodei, and Hassabis pitched world leaders on an US-led international AI governance body, chip trade restrictions on China, and controlled access to frontier models.
• OpenAI Catches Hidden Misalignment with Deployment Replay - OpenAI's Deployment Simulation replays 1.3M real user conversations through candidate models and found a novel reward-hacking bug in GPT-5.1 before release.
• Baseten Closes $1.5B Round as Open-Source AI Matures - Baseten's raise at a $13B valuation signals that open-source models are displacing closed APIs in enterprise AI.
• Odyssey Raises $310M as Amazon Bets on World Models - Amazon leads a $310M round into Odyssey, a startup building world models that simulate physics rather than just language.
• Facebook AI Mode Mines Public Posts With No Opt-Out - Meta launched AI Mode on Facebook, turning billions of public posts into AI-synthesized search answers with no opt-out for 2 billion daily users.
• Behavox Raises $175M From BlackRock for AI Compliance - Behavox lands preferred equity from BlackRock's HPS as compliance demand rises ahead of Colorado and EU AI Act deadlines.
• Microsoft Bets on DeepSeek V4 to Cut Copilot Costs - Microsoft is shifting Copilot Cowork to usage-based pricing and testing DeepSeek V4 on Azure, trading 57x cheaper tokens for ongoing geopolitical risk.
• Amazon Eyes Third-Party Trainium Sales to Rival Nvidia - Andy Jassy hints Amazon will sell Trainium3 racks to outside data centers, citing a potential $50B revenue run rate and sold-out chip supply.
• US Confronts ASML Over Suspected EUV Exports to China - Commerce Secretary Howard Lutnick told ASML executives the Trump administration believes EUV-related gear may have reached China in violation of export controls; ASML denies the claim.
• Canada's Pension Giant Bets C$1B on India AI Data Centers - CPP Investments commits C$1 billion to Indian data center operator CtrlS in a dual equity-and-JV structure.
• AI Label Backlash - 60% of US Consumers Are Turned Off - Two surveys from WordPress VIP and Gartner find a majority of US consumers are put off by AI labels in brand messaging, even as companies race to build AI visibility for search engines.
• Alibaba's Qwen-Robot Suite Targets Physical AI Work - Alibaba launches three open-weight models for robot manipulation, navigation, and world prediction, built on a shared Qwen3.5 backbone.
• Pramaana Labs Raises $27M for Provably Correct AI - Pramaana Labs uses the LEAN proof language to attach a mathematical certificate to every AI answer in high-stakes domains like tax, law, and drug discovery.

Reviews

• DiffusionGemma 26B Review: 4x Faster, Real Tradeoffs - Google DeepMind's DiffusionGemma generates 1,000+ tokens per second through parallel diffusion, trading 5-19 benchmark points against Gemma 4 for speed.
• Mistral Medium 3.5 Review: Open Agent, Sharp Teeth - Mistral's 128B open-weight model consolidates reasoning, coding, and vision into one checkpoint, with remote agents that file pull requests autonomously.

Guides

• AI for Small Business Marketing - A Beginner's Guide - Step-by-step guidance for using AI in content creation, email, social media, and ads - no marketing background required.
• How to Use AI for Home Improvement Projects - A practical walkthrough for planning, designing, and budgeting home improvement work with AI tools.

Tools

• Best AI Synthetic Data Tools 2026 - Ranked - The top synthetic data tools ranked by price, quality, and use case - from open-source SDV to Tonic.ai Fabricate and K2view.

Science

• Emergent Alignment, Agent Memory, and Smarter Reasoning - Three arXiv papers: a conscience mechanism for ethical training, shared memory for agent populations, and selective verification that cuts test-time compute waste.
• Faster Agents, Skewed Evals, and Brand Bias in LLMs - Three papers: agents compiled into 8-13x faster state machines, benchmark scores that shift with compute budget, and big brands monopolizing LLM recommendations.
• AI Engrams, Cognitive Debt, and Agent Trust - Three papers on what lives inside a trained model, how AI dependence erodes human cognition, and whether AI teams can calibrate trust.

Models

• GPT-5.1 - OpenAI's coding and agentic flagship with 400K context, configurable reasoning effort, and 76.3% on SWE-bench Verified.
• Voxtral TTS - Mistral's first open-weight text-to-speech model: 4B parameters, 70ms latency, voice cloning from 3 seconds of audio.
• MAI-Transcribe-1.5 - Microsoft's second-generation speech-to-text model with 2.4% WER, 43-language support, and 5x faster long-audio processing.
• Qwen3.7-Plus - Alibaba's first multimodal agent model, combining GUI grounding, 1M-token context, and text-plus-vision input at $0.40/M tokens.
• Qwen-RobotManip - Alibaba's generalist VLA model for robotic manipulation, trained on 38,100+ hours of open-source data and ranked first on the RoboChallenge generalist track.

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