Awesome Agents Weekly

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

Today closed GitHub Copilot's first full month under token-based billing, and the bills landing in developer inboxes are settling the debate about what agentic AI actually costs. OpenAI shipped its first custom chip, five intelligence agencies warned that AI-enabled cyberattacks are months away, and DeepMind published a roadmap from AGI to superintelligence. The gap between "AI is coming" and "AI is here" keeps shrinking.

Pick of the Week

OpenAI Ships Jalapeño - Its First Custom AI Chip

For five years, OpenAI has been buying time on the same silicon everyone else rents. That changed on June 24 when Broadcom CEO Hock Tan walked into OpenAI's offices carrying a 300mm wafer holding roughly 50 Jalapeño ASICs - built on TSMC's 3nm process - and handed it to Sam Altman. The chip uses a systolic array with eight HBM stacks and targets roughly 50% lower cost per inference token compared to current GPU-based alternatives. It's the first step toward OpenAI owning its own cost structure rather than renting it from Nvidia indefinitely.

This Week on Awesome Agents

News

• GitHub Copilot Goes Metered - Agentic Devs See 25x Bills - The first month under token-based billing closed today, with agentic developers projecting $600-$1,200/month per developer - up from a flat $39.
• Colorado's AI Law Takes Effect Today, Already Gutted - Once framed as America's first comprehensive AI consumer protection law, SB 24-205 became technically enforceable today with its core protections stripped by xAI's federal lawsuit, DOJ intervention, and a replacement bill signed in May.
• Five Eyes Warn: Frontier AI Will Break Defenses in Months - Six intelligence directors from the US, UK, Australia, Canada, and New Zealand issued a joint warning that the window to prepare for AI-enabled cyberattacks is measured in months, not years.
• Anthropic Tells Senate Alibaba Stole Claude's Capabilities - Anthropic accused Alibaba's Qwen lab of running 28.8 million unauthorized exchanges against Claude through 25,000 fake accounts between April and June, calling it the largest known distillation attack against an AI model.
• DeepMind Maps Four Routes from AGI to Superintelligence - A 57-page paper co-authored by DeepMind co-founder Shane Legg identifies four pathways from AGI to ASI and six bottlenecks that could block each route.
• White House Forces GPT-5.6 Into a Staged Rollout - The Trump administration requires individual government approval for every GPT-5.6 customer, citing cybersecurity capabilities that rival Anthropic's restricted Mythos model.
• Meta Restricts Claude Code Over Training Data Leakage - Meta stopped its engineers from using Claude Code and Codex, citing distillation risk - an AI lab building its own frontier models can't afford its codebase appearing in a rival's training data.
• US Clears Anthropic Mythos 5 for 100 Trusted Partners - Commerce partially lifted its two-week export ban, clearing Mythos 5 for roughly 100 US companies while keeping Fable 5 blocked with no timeline.
• Ban on Anthropic Models Fuels Asia's AI Race - Two weeks after the US barred global distribution of Mythos 5 and Fable 5, Tokyo's Sakana AI and China's 360 Security announced frontier-class alternatives timed to fill the gap.
• AI Job Cuts in 2026 Already Beat All of Last Year - US employers cited AI in 87,714 job cuts through May 2026, already past the full-year 2025 total of 54,836, while companies posting AI-driven revenue growth simultaneously report record profits.
• Google Loses Four AI Stars to Rivals in Six Days - Noam Shazeer, John Jumper, and two more Gemini researchers left for OpenAI and Anthropic between June 18 and 24, sending Alphabet shares down nearly 10% from their May peak.
• Google Delays Gemini 3.5 Pro Past Its I/O Promise - Sundar Pichai told developers to expect Gemini 3.5 Pro in June; it missed the window and slipped to July as four senior researchers departed.
• Qualcomm Bets $14B to Shatter Nvidia's AI Lock-In - Qualcomm confirmed a $3.9B acquisition of AI software startup Modular and advanced talks to buy RISC-V chip designer Tenstorrent for up to $10B, betting that a portable compiler can dislodge CUDA from the data center.
• Ford Rehires 350 Engineers After AI Quality Fails - Ford won the 2026 JD Power quality study - its first mainstream brand crown in 16 years - by reversing its automation-first strategy and bringing back 350 veteran engineers to fix what the machines got wrong.
• AI Patched Firefox Before Pwn2Own - OpenAI's Security Pivot - OpenAI's GPT-5.5-Cyber found CVE-2026-8390 in Firefox's WebAssembly engine before Pwn2Own Berlin, causing five of six registered exploit teams to withdraw.
• Reflection AI Locks $6.3B Compute Deal With SpaceX - The AlphaGo co-founder's open-source AI lab will pay $150M monthly for Nvidia GB300 chips at Colossus 2, committing $6.3 billion total before shipping a single public model.
• Agility Robotics Goes Public in $2.5B SPAC Deal - Agility merged with Churchill Capital Corp XI at a $2.5B valuation, becoming the first publicly traded pure-play humanoid robot company with robots already earning commercial revenue.
• California Signs Anthropic Deal for Claude at Half Price - Newsom gave all state agencies access to Claude at 50% off, publicly breaking from the federal government's stance toward Anthropic.
• Apple's Vision Pro Chief Joins OpenAI Hardware Unit - Paul Meade, Apple's VP for the Vision Pro and smart glasses, is joining OpenAI's io hardware division - the most senior hardware engineering defection yet in a two-year talent war.
• Anthropic Locks In Micron as HBM Defines the AI Race - Micron signed a multi-year HBM supply deal with Anthropic and took an equity stake in its Series H, as memory bandwidth becomes the rate-limiting constraint for frontier AI.

Reviews

• GLM-5.2 Review: Best Open-Weight Coder at 1/6 Cost - Z.ai's GLM-5.2 delivers frontier coding performance with open weights and an MIT license at roughly one-sixth the cost of GPT-5.5.
• Grok 4.3 Review: xAI Bets on Price Over Prestige - Grok 4.3 cuts prices by up to 83% and adds native video input, carving out a credible spot as the most cost-efficient frontier model - with real caveats on coding and latency.

Guides

• How to Use AI for Time Management - A Beginner's Guide - A practical guide to using Reclaim, Motion, and ChatGPT to schedule your day, protect deep work time, and cut inbox overhead.
• How to Use AI for Home Buying - A Beginner's Guide - A practical walkthrough for first-time and repeat buyers on using AI tools to research neighborhoods, compare mortgages, analyze inspection reports, and prep for negotiation.

Tools

• Best LLM Gateways 2026 - LiteLLM, Portkey, and 5 More - A hands-on comparison of seven LLM gateway and routing tools: LiteLLM, Portkey, Helicone, OpenRouter, Martian, Cloudflare AI Gateway, and Bifrost.

Science

• Tandem Training, World Models, and Efficient Agents - Three papers on making RL reasoning legible across models, fixing broken world model latent states, and training small agents to beat their teachers.
• Refusal Gaps, Prompt Bleed, and Scaling's Logic Limit - New research shows LLM safety depends heavily on persona training, that prompt modules interfere in deployed agents, and that scaling alone can't reach symbolic reasoning.
• Quantization's Hidden Tax, Cliff Tokens, Smarter Memory - Three papers reveal hidden costs in quantized reasoning models, single-token failure triggers, and a new memory framework that cuts agent memory errors by up to 79%.
• AI Diagnosis, Cache Efficiency, and Agent Security - A 32B medical model beats DeepSeek-R1 in rare disease diagnosis, a KV cache method keeps 97% accuracy at 3% memory use, and a new benchmark red-teams agentic AI systems.
• AI Research: Orchestration Beats Scale, Small Models Win - Sakana Fugu tops SWE-Bench Pro by routing tasks across rival LLMs, Microsoft's 9B browser agent beats OpenAI Operator, and a 3B model matches DeepSeek V3.2 on math.

Models

• Grok 4.5 - xAI's 1.5-trillion-parameter V9 model in private beta at SpaceX and Tesla, with early internal evals claiming performance near Claude Opus 4.8.
• Gemini 3.5 Pro - Google DeepMind's upcoming flagship with a 2M-token context window and Deep Think reasoning, announced at I/O and expected in July.
• Claude Mythos 5 - Anthropic's restricted Mythos family, now cleared for roughly 100 US trusted partners: same weights as Fable 5, without safety classifiers for cybersecurity and biology.
• North Mini Code - Cohere's first developer-focused model - 30B sparse MoE with 3B active parameters, free Apache 2.0 license, 256K context, and 33.4 on the AA Coding Index.
• Sakana Fugu - Sakana AI's orchestrator that coordinates Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro to beat each of them individually on SWE-Bench Pro and GPQA-Diamond.
• GPT-5.5-Cyber - OpenAI's cybersecurity-specialized fine-tune of GPT-5.5, restricted to vetted defenders and rated 85.6% on the CyberGym benchmark.
• Fara-1.5 - Microsoft Research's open-weight browser computer use agent family (4B, 9B, 27B) that beats OpenAI Operator and Gemini 2.5 Computer Use on Online-Mind2Web.
• VibeThinker-3B - WeiboAI's 3B dense reasoning model fine-tuned from Qwen2.5-Coder-3B, posting AIME 2026 scores that match DeepSeek V3.2 (671B).
• ERNIE 5.1 - Baidu's 800B-parameter text-focused MoE model claiming the top Chinese model slot on LMArena, built at 6% of comparable training costs.
• Sora 2 - OpenAI's physics-accurate video generation model with synchronized audio, available API-only until its September 24, 2026 sunset.
• Dreamina Seedance 2.0 - ByteDance's top-ranked AI video model with native joint audio-video synthesis, multi-shot support, and up to 12 multimodal reference inputs per generation.
• Kling 3.0 - Kuaishou's first commercially available AI video model to ship native 4K at 60fps, with multilingual audio, multi-shot storyboarding, and a $0.075/s API.
• Runway Gen-4.5 - Runway's Autoregressive-to-Diffusion video model, which led the Artificial Analysis Elo rankings at launch before Seedance 2.0 and Kling 3.0 surpassed it.
• Wan 2.7 - Alibaba's open-source video generation model with MoE architecture, native audio, first-and-last-frame control, and 1080p output up to 15 seconds.
• Grok Imagine Video 1.5 - xAI's top-ranked image-to-video model on Artificial Analysis, producing 720p clips with native audio at $0.14/s - 86% cheaper than Sora 2 Pro.
• SkyReels V4 - Skywork AI's unified multi-modal video model creating 1080p/32FPS video and synchronized audio from a single dual-stream diffusion transformer.
• HappyHorse-1.0 - Alibaba's 15B-parameter video generation model that ranked #1 on Artificial Analysis, producing 720p-1080p clips with joint audio-video synthesis in a single forward pass.

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