Awesome Agents Weekly

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

Over seven days, Anthropic closed a $65 billion round, Cognition closed a $1 billion round, SoftBank committed €75 billion to French AI data centers, and ByteDance revealed it may spend $70 billion on AI this year alone. These numbers build up into a straightforward observation: the infrastructure race has moved to a scale where competitors without sovereign backing or major enterprise contracts are going to struggle to stay in it. Meanwhile, the agents that all this capital is funding started spending real money this week - on Robinhood, 27.5 million retail customers can now let Claude or ChatGPT manage their stock portfolios while regulators work out who's responsible when it goes wrong.

Pick of the Week

Anthropic Closes $65B Series H at $965B Valuation

Anthropic went from $380 billion to $965 billion in 105 days, closing the largest private AI funding round in history and becoming the most valuable private AI company - pushing past OpenAI's $852 billion mark from two months ago. The number that matters most isn't the valuation; it's the $47 billion in run-rate revenue underpinning it. This isn't a speculative raise by a company betting on future usage - it's a company converting present usage into capital at a rate that makes IPO logistics feel closer than they did six months ago. The harder question for the second half of 2026 is whether Anthropic can hold its safety-first positioning once it becomes a public company and quarterly revenue growth is the primary accountability structure.

This Week on Awesome Agents

News

• Anthropic Closes $65B Series H at $965B Valuation - The largest private AI funding round in history, backed by $47B in run-rate revenue, makes Anthropic the most valuable private AI company ahead of OpenAI.
• Cognition Raises $1B at $25B as Devin Hits $492M ARR - More striking than the $25B valuation: over 90% of Cognition's own codebase is now written by the AI agent it sells to enterprise customers.
• SoftBank Commits €75B to French AI Data Centers - A phased €75B commitment targeting 5 GW of capacity in Hauts-de-France by 2031, which would create the largest AI compute cluster in Europe.
• ByteDance Plans $70B AI Capex, Tripling Last Year - ByteDance is weighing up to $70B in AI capital expenditure for 2026, nearly tripling its 2025 spend, while locking in offshore compute and a Qualcomm chip deal to work around US export controls.
• BadHost: The Auth Bypass Lurking in 325M AI Systems - CVE-2026-48710 lets a single malformed HTTP header bypass authentication across vLLM, LiteLLM, FastAPI, and every MCP server built on Starlette - patch available since May 31.
• Anthropic Ships Opus 4.8 with Multi-Agent Workflows - Opus 4.8 scores 69.2% on SWE-bench Pro, up from 4.7's 64.3%, adds parallel subagent orchestration in research preview, and keeps pricing unchanged at $5/$25 per million tokens.
• Microsoft Launches Polaris and Foundry Local at Build 2026 - Project Polaris replaces GPT-4 Turbo in GitHub Copilot by August using Microsoft's Maia accelerators; Foundry Local ships as a ~20 MB on-device runtime with no cloud dependency.
• Nvidia Enters the PC Market With RTX Spark Superchip - Nvidia's first Windows SoC pairs a 20-core Arm CPU with a Blackwell 2.0 GPU on TSMC 3nm, launching in 30+ laptop designs this fall and challenging Apple Silicon directly.
• Nvidia Cosmos 3 Is the First Open Physical AI Model - Trained on 20 trillion tokens, Cosmos 3 handles text, images, video, sound, and robot action data in one forward pass, fully open-weight at 16B and 64B sizes.
• GitHub Copilot Goes Token-Based: Devs Report 25x Bills - Copilot switched to token billing on June 1; some developers report their $29/month plans becoming $750 as agentic workflows drain credits in a single session.
• Robinhood Opens AI Agent Trading to 27M Retail Users - Robinhood launched MCP-powered agentic trading in beta, letting Claude and ChatGPT manage stock portfolios for retail customers ahead of any regulatory framework for it.
• Gemini CLI Dies June 18 - Google Goes Closed-Source - Google is retiring free access to its Apache 2.0 Gemini CLI on June 18, replacing it with the closed-source Antigravity CLI after accepting over 6,000 community pull requests.
• OpenAI Governance Doc Targets California and EU AI Law - OpenAI published its first compliance framework mapping to California SB 53 and the EU AI Act, while critics note its Preparedness Framework quietly dropped manipulation from risk categories last April.
• CNN Sues Perplexity in First TV Network Copyright Case - CNN filed a federal copyright and trademark suit, becoming the first television network to take legal action against an AI search company over scraping more than 17,000 stories.
• DuckDuckGo Traffic Triples After Google's AI Search Pivot - Nine days after Google I/O 2026 made AI Overviews mandatory with no opt-out, traffic to DuckDuckGo's no-AI search page tripled.
• YouTube Takes AI Video Labeling Into Its Own Hands - YouTube now auto-detects and labels photorealistic AI video without creator disclosure, using C2PA metadata and internal signals in a shift that removes creator opt-in from the equation.
• IRGC Hackers Used AI to Build Malware During Iran War - Iranian IRGC-linked group Nimbus Manticore used AI coding tools to accelerate development of a new backdoor in real time across three campaign waves during Operation Epic Fury.
• China Locks Down AI Talent at Alibaba and DeepSeek - China extended travel restrictions to AI researchers at private firms including Alibaba and DeepSeek, requiring government approval before they can leave the country.
• Groq Raises $650M to Pivot From Chip Maker to Cloud - After licensing chip technology to Nvidia for $20B, Groq is raising $650M from existing investors to rebuild itself as an AI inference cloud provider.
• IBM and Red Hat Bet $5B on AI to Secure Open Source - Project Lightwell deploys 20,000 engineers and AI to patch open source vulnerabilities against exact deployed versions without forcing upgrades.
• Mistral Physics AI Shrinks Days of Simulation to Seconds - Mistral bought Vienna-based Emmi AI and launched Physics AI - models that replace multi-day engineering simulations with seconds of inference on a single GPU.
• Mistral Vibe Adds Work Mode and a VS Code Extension - Mistral rebranded Le Chat as Vibe and shipped a VS Code extension backed by Mistral Medium 3.5 at 77.6% SWE-Bench Verified, making it a direct competitor to GitHub Copilot and Cursor.
• Visa Bets on Replit to Win Agentic Payments Race - Visa took a stake in Replit and integrated its Trusted Agent Protocol, placing payment identity infrastructure directly inside the tools developers use to build AI agents.

Reviews

• Antigravity 2.0 Review: Agent-First, Rocky Launch - Google's Antigravity 2.0 redesigns the platform as a five-surface agent suite; the architecture is ambitious, but the launch was a mess.
• Kore.ai Artemis Review: Enterprise Agent Control Plane - Artemis's compiled blueprint language and governance-first architecture earn a 7.8/10, making it the most governance-serious agent platform available right now - though it's Azure-only until October 2026.

Guides

• How to Use AI for Resume and Interview Prep - A practical guide to using AI tools to write stronger resumes, craft tailored cover letters, and prepare for job interviews without letting the AI do all the work.
• How to Use AI to Summarize Long Documents and PDFs - A step-by-step walkthrough for uploading PDFs into ChatGPT, Claude, and Gemini and writing prompts that produce useful, trustworthy summaries.

Leaderboards

• AI Image Generation Leaderboard: Best Models 2026 - Current rankings across GPT Image 2, Recraft V4.1, HiDream-O1-Image, FLUX 2, Midjourney v8.1, and Ideogram 3.0, scored on human preference, text rendering, and photorealism.

Science

• Cut CoT Costs, Fix Agent Memory, Test Clinical AI - SLAT trims reasoning length by 50% without accuracy loss; AdaCoM rescues frozen agents on long-horizon tasks; EHRBench's 960K clinical QA items show consistent LLM gaps in diagnosis and prognosis.
• Reasoning Capitulation, Faster Guardrails, Curation Risk - Reasoning models' chains stay correct but final answers flip under adversarial pressure; latent-space guardrails run 12.9x faster; and adding human curation can invert alignment improvements in multi-model training loops.
• Alignment Faking, Agent Collusion, and Brittle Safety - Three papers decompose alignment faking into measurable drivers, show safety-aligned agents will collude when strategic advantage is available, and find standard guardrails miss the worst context-flip failures.

Models

• NVIDIA Cosmos 3 - Nvidia's first fully open physical AI omnimodel with Mixture-of-Transformers architecture at 16B and 64B sizes, natively handling text, images, video, sound, and robot actions in one pass.
• Claude Opus 4.8 - Anthropic's May 2026 flagship scores 69.2% on SWE-bench Pro, ships dynamic parallel workflows, and maintains a 1M-token context window at $5/$25 per million tokens.
• Qwen3.7-Max - Alibaba's agent-first flagship tops Terminal-Bench 2.0 and SWE-Bench Pro at roughly one-sixth the cost of Claude Opus 4.7, with a 1M-token context window.
• NVIDIA SANA-WM - A 2.6B-parameter world model that creates 60-second 720p camera-controlled video on a single H100, open-weight and built for robotics simulation.

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