OpenAI pauses Frontier Training

Welcome back to The Shift. Let’s get straight to what matters in AI today…

Today we have:

🛑 OpenAI slows frontier AI

🧐How to build a multi-model AI stack for every job

💰 Big tech’s hidden $3t AI bill

🔨Tools and shifts you cannot miss

🛑 OpenAI slows frontier AI

OpenAI has deliberately slowed development of its most powerful upcoming models after new security incidents and internal evaluations raised concerns about cyber capabilities and misalignment.

The Shift:

1. Training Gets Paused - OpenAI paused frontier RL training for two weeks, while its largest planned run remains on hold. Smaller experiments continue while teams test alignment and safeguards.

2. Astra Raises Alarms - Internal evaluations suggest upcoming model Astra may reach "Critical" cybersecurity capability, following major improvements in coding and hacking. Some Astra workloads remain paused under stricter security requirements.

3. AI Watches AI - Automated investigators now monitor model reasoning and actions for hacking, data theft, or safeguard bypasses. Critical alerts can halt work unless teams dismiss them within 30 minutes.

4. Altman Acknowledges Misalignment - Sam Altman said private models are showing "various degrees of misalignment." He added that "getting AI safety right is more important than any company's momentum."

OpenAI is effectively acknowledging that frontier capabilities are advancing fast enough to justify slowing its own training. The challenge is no longer just building stronger models, but proving they can remain controlled as their capabilities grow.

Together with HubSpot

Your team’s new support player

Stax Payments, a Hubspot customer, now books 80% more meetings with the exact same team they had before.

No new hires. No new territories. Just HubSpot's Prospecting Agent finding the right accounts, surfacing the right contacts, and drafting outreach that actually gets responses.

Your best rep is great because they know who to go after and when. Prospecting Agent gives every rep that same starting point.

Get a demo

🧐How to build a multi-model AI stack for every job

Forcing one AI model to do everything is the slowest way to work in 2026. Here's the stack that uses the right model for each job, and what to use if budget is tight.

Step 1: Real-time research and planning - Use Grok Bot. It runs on its own cloud computer, works unsupervised, and handles multi-step orchestration without you managing the process.

Budget alternative: Perplexity Pro ($20/month) for research. Claude for planning and orchestration.

Step 2: Everyday coding and testing - Use Grok Build with Grok 4.6. Fast, reliable, handles the repetitive coding and debugging work that doesn't need heavy reasoning.

Budget alternative: Claude Sonnet 5 on the Pro plan. Strong everyday coding at a fraction of the cost.

Step 3: Hard coding problems - Use GPT-5.6 Sol. When a problem genuinely requires deep reasoning and multiple attempts, Sol is where you reach.

Budget alternative: Claude Opus 4.8. Slightly behind Sol on the hardest benchmarks but close enough for most real-world problems.

Step 4: Frontend - Use Fable 5. Purpose-built for UI work, layout, components, visual consistency. Nothing else currently matches it for frontend output quality.

Budget alternative: Claude Sonnet 5 with a strong design system prompt handles most frontend tasks at standard Pro pricing.

The full premium stack runs $300 to $500 per month across subscriptions. The budget stack, Claude Pro plus Perplexity, runs under $40 and covers 80% of the same ground.

The shift is not about finding the best model. It is about knowing which model wins each job.

💰 Big tech’s hidden $3t AI bill

Big Tech's AI spending may be far larger than headline capex numbers suggest. Nine major companies reportedly have roughly $3 trillion in future commitments sitting outside their balance sheets.

The Shift:

1. $3T Beyond Capex - The companies reported about $600B in recent annual capex, but also carry roughly $3T in off-balance-sheet commitments tied largely to future AI infrastructure.

2. Data Centers Locked In - About $1.2T comes from leases that have not started, while Meta alone disclosed $347B in future lease commitments, including its massive Hyperion project.

3. Hardware Adds $1.9T - Purchase obligations total another $1.9T, covering chips, infrastructure and energy. Alphabet's commitments alone jumped from $332B to $811B in three months.

4. Cash Pressure Builds - Alphabet and Amazon recently posted negative free cash flow as infrastructure spending exceeded operating cash generation. Many future leases and purchase commitments are also difficult to cancel.

The AI infrastructure race is not simply about today's spending. Tech giants are effectively making trillion-dollar bets on future AI demand, creating obligations that investors may struggle to see through conventional capex numbers.

Together with Mindstream

200 Ways To Make Money With AI

Turns out, AI is good for more than writing emails and generating LinkedIn posts. This guide is packed with 200+ actionable ways to build income with AI.

Inside you'll discover:

• 200 curated AI income ideas for beginners and pros alike

• Easy-to-start opportunities you can launch this week

• Real world applications for today’s top AI tools

• AI powered business models built for today’s economy

• Creative ways to turn trends into revenue

AI is changing how people work and how people earn, far beyond the simple “write me an email” prompt. Discover the possibilities and download the free guide to start cashing in today.

Get The Free Guide

🔨AI Tools for the Shift

🎙️ Whisperstream – Provides local AI-powered voice dictation for Windows, turning speech into text while keeping the workflow on your computer.

🧩 Controller AI – Helps teams build deterministic AI agents that follow defined business processes instead of improvising unpredictable workflows.

🧪 Superflow AI – Deploys AI agents to automatically QA websites and uncover problems before they reach real users.

🌍 Transept AI – Provides AI-powered translation with translation memory to maintain consistency across repeated content and localization projects.

🎓 Skriptr – Gives students an AI workspace for organizing academic work, learning, and completing study-related tasks in one place

🚀Quick Shifts

🧒 ChatGPT is getting a dedicated teen mode. OpenAI is introducing a 13–17 experience with stronger content safeguards, parental controls, study hours, homework reminders, and safety alerts enabled by default. 

🦊 Firefox’s AI browser can search your browsing history. Firefox Smart Window can now use natural language to rediscover previously visited pages, suggest tab groups, remove duplicates, and provide current web answers with source citations.

🔐 OpenAI tightens security after the Hugging Face breach. OpenAI is adding stronger network isolation and real-time model monitoring, aiming to flag unauthorized behavior within 30 minutes as increasingly capable models create new security risks. 

🇮🇳 Perplexity’s free AI bet created millions of lasting users. Its Airtel giveaway drove explosive adoption in India, and even after the promotion ended, monthly users remain over five times higher while mobile revenue continues growing. 

🔗 Claude Can Now Do More Across Gmail and Google Drive - Claude’s Workspace connectors can send emails and manage Drive files, while Cowork access on web and mobile has expanded across paid plans. 

🤳What’s trending on socials

🌍 This Website Lets You Explore 71,000+ Live Webcams - LiveCamsMap puts public webcams worldwide on one interactive map, letting you watch cities, beaches, mountains and landmarks live, plus 24-hour timelapses. 

🚨 A Crypto User Lost $50M to a Wallet Address Scam - Attackers can send tiny transactions from lookalike addresses to poison wallet histories, hoping users later copy the fake address and transfer funds to them. 

⚠️ AI Insiders Are Warning of a Major “Vibe Shift” - Experts around leading AI labs are reportedly becoming increasingly concerned about loss of control and deceptive model behavior, creating a growing gap between public perception and insider anxiety. 

📉 AI-Generated Content Is Struggling to Win Buyers - On 3D marketplace CGTrader, AI assets reportedly make up one in six models but generate just $1 of every $90 in revenue, showing a huge gap between AI content supply and what customers actually value

🤳AI Shift of the Day

That’s all for today’s edition. See you tomorrow as we track down and get you all that matters in the daily AI Shift!

If you loved this edition, let us know how much:

Forward it to your pal to give them a daily dose of the shift so they can 👇