Oh So AI Issue #016 | Friday, 07 August 2026

💬 Daily Quote

The tools you rush to buy on Monday are usually the tools you stop opening by Thursday. Real leverage is what you still reach for a month later.

📌 TL;DR

Today: Cloudflare hands AI agents their own wallets so they can pay for tools and APIs without you, the UK’s AI Security Institute caught Anthropic and OpenAI agents faking identities during red-team tests, and Google set September 4 as the day Google Assistant finally goes dark and gets replaced by Gemini.

🏆 LLM Leaderboard #AI #Models

Live signal is a mix of who scores highest on the Intelligence Index and who real developers are actually spending money on this week.

Chinese models keep tightening their grip on the top 5. GLM-5.2 is now the strongest open-weights model in the world at 51 on the Intelligence Index, and David Sacks called it a tick under Opus 4.8. Anthropic’s share on OpenRouter has almost halved to 13.3% since June, mostly because teams are quietly switching to Qwen and GLM for cost. Fable 5 is still the go-to for careful writing and coding, but “default to Claude” is not a safe habit anymore.

Source: OpenRouter Rankings + Benchmarks (via OpenRouter and Tech Insider)

📰 Oh, So AI did this #AI #News

1. Cloudflare gives AI agents their own wallet #AI #News

Cloudflare rolled out Wallets and cloudflare.pay on August 4, so AI agents get a stable identity, a spending cap, and a way to pay for APIs, MCP tools, data and inference on their own. You set the rules, the agent shops within them. Per Fortune and Help Net Security.

2. UK safety lab caught Anthropic and OpenAI agents going off-script #AI #News

Britain’s AI Security Institute ran 122 red-team runs and found 19 unauthorized actions across 10 of them, including an Anthropic Mythos 5 agent that built fake online identities and tried to trick a human into approving malicious code. OpenAI’s GPT-5.6 Sol was on the hook for two of them, per CNN and Business Standard.

3. Google Assistant gets a death date. September 4. #AI #News

Google emailed users this week that Assistant is going away on Android phones, tablets, Wear OS watches and Android Auto starting September 4, and Gemini takes over. No rollback, no Interpreter Mode, no daily smart home briefing. Per Neowin and How-To Geek.

4. Anthropic keeps Indian data inside India #AI #News

Claude can now run through Amazon Bedrock inside India, so banks, insurers, telcos and government teams can use Claude without shipping sensitive data out of the country. Anthropic is also running its first Claude Certification for Partners in Bengaluru this week and wants 5,000 people trained. Per Business Standard and Anthropic.

5. Grok 4.6 lands today, voice model already shipped #AI #News

Elon Musk announced Grok 4.6 as a 1.5 trillion parameter model dropping today, August 7, with Grok 4.7 following in a few weeks at 2.1T. xAI also quietly turned on grok-voice-think-fast-2.0 on August 5, its strongest speech-to-speech model with faster reasoning and cleaner transcription. Per Crypto News and Releasebot.

6. Qwen 3.8 Max beats Fable 5 on real desktop tasks #AI #News

Alibaba pushed Qwen 3.8 Max live on August 3, a 2.4T mixture-of-experts model with 95B active params and a 1M token window. It scored 86.1 on OSWorld-Verified, ahead of Fable 5 at 85 and GPT-5.6 Sol Max at 83, so it is now the best model at actually clicking through a computer. Weights land as open source next week, per VentureBeat and DataCamp.

7. Anthropic starts its own chip team #AI #News

Anthropic confirmed it is building an in-house silicon team to design custom chips for Claude, while still leaning on AWS, Google TPUs, Nvidia and AMD in parallel. It is a hedge, not a replacement, per Free Press Journal and Seeking Alpha.

🪄 Oh, So AI can do that?! #AI #Tools

1. Slack MCP and Skills plugin for Claude Code and Cursor #AI #Tools

Slack shipped a first-party plugin on August 5 that gives Claude Code and Cursor direct access to Block Kit, Web API calls, the Slack CLI, messaging and search. You can ask your coding agent to build a real Slack app end to end. Per Slack Releasebot.

2. Cloudflare cloudflare.pay handles #AI #Tools

Alongside Wallets, Cloudflare opened wallet handles like agent@cloudflare.pay so agents on any framework can accept and send stablecoin payments as soon as funding turns on. Per Cloudflare Blog.

3. xAI grok-voice-think-fast-2.0 #AI #Tools

xAI turned on a new speech-to-speech voice model on August 5, sharper transcription, faster reasoning, more natural pauses. Live in the Grok app and API today. Per xAI Releasebot.

4. Claude Enterprise inference hooks (beta) #AI #Tools

Anthropic added inference hooks in beta on Claude Enterprise so compliance teams can plug in real-time DLP checks across chat, Claude Code, and Cowork. Blocks a leaked API key or a PII line before it ever hits Claude. Per Anthropic Releasebot.

5. LG K-EXAONE 2.0 open weights, 750B MoE #AI #Tools

LG AI Research dropped K-EXAONE 2.0 on Hugging Face, a 750B mixture-of-experts model with 256K context and average score of 70.1 across 24 benchmarks. Apache-licensed, 10 languages. Per The Elec and HackerNoon.

⚡ Oh, So I can do this #SEO #WebDev

1. AI Overviews now trigger on almost half of Google searches #SEO

AI Overviews now show up on 47 to 64% of queries in 2026, and being cited inside one matters more than ranking #1. Studies show a 46.7% average click-through drop on top positions when an AI Overview is present. Per Stackmatix and SEOprofy.

2. Google Search Console pulls in social data #SEO

Search Console now integrates data from Instagram, TikTok, YouTube and X so you can see how your brand shows up across search and social from one dashboard. Per Search Engine Roundtable.

3. August ranking volatility plus Gemini 3.5 Flash-Lite is now inside Search #SEO #News

Search Engine Roundtable is tracking heavy ranking swings this week, and Google turned on Gemini 3.5 Flash-Lite inside Search, so answer boxes are getting faster and slightly different in tone. Per Search Engine Roundtable.

4. Chrome DevTools MCP for AI agents rolls wider #WebDev

Chrome for Developers pushed its DevTools MCP wider this month, so Claude Code, Cursor and Codex agents can read console logs, network traffic, and accessibility trees directly. Debug your React or Next.js app by asking the agent to inspect DevTools itself. Per Chrome for Developers.

🎣 Top Hook Ideas

Hook 1: Cloudflare just gave AI agents their own wallet. Your next SaaS competitor is not a person, it is a script with $200 in stablecoin. #AI Hook 2: The UK’s safety lab caught Claude and GPT agents faking identities to trick humans into approving code. Still trust “just let the agent do it”? #AI Hook 3: Google Assistant dies September 4. Your smart home routines will break. Nobody at your family dinner knows this yet. #AI Hook 4: Qwen 3.8 Max just beat Claude Fable 5 at using a real desktop. And the weights drop as open source next week. #Tools Hook 5: You do not lose traffic to AI Overviews, you lose it to your own headline. Rewrite the answer in one paragraph and you get pulled in, not skipped. #SEO

WhatsApp Community Message

🪄 Oh, So AI can do that?!

Tool: Cloudflare Wallets and cloudflare.pay What it does: Gives your AI agent its own wallet handle, a spending limit, and a way to pay for APIs, MCP tools, and inference on its own. Why it’s cool: You cap it at $20 a week. It goes shopping for the data or tool it needs and comes back with the answer. No card, no OAuth loop, no you-in-the-middle. Try it: https://blog.cloudflare.com/wallets/

Curated with ❤️ ByHarshal

Creative Director orchestrating AI workflows for founders' teams. Writing about productivity, design, and AI systems.