OpenAI’s $12B ARR 💰, Meta’s Personal Superintelligence 🧠, Linear’s Agent Guidelines 🤖

OpenAI has reached $12 billion in annualized revenue, roughly doubling its revenue in the first seven months of the year ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR Together With TLDR AI 2025-07-31 Software used to be about features. With AI, it's about real value delivered (Sponsor) When you buy an Office license, you're paying for features or capabilities. With AI, features and capabilities are nearly unlimited, and the value of software becomes what it actually does. That's why OpenAI and Anthropic charge by tokens used rather than seats. The change is inevitable, but many monetization models are stuck in the past. Read more about the three eras of software value—and the cost of not keeping up with the current one—on the Metronome blog. P.S. Later this month, Metronome will release a full whitepaper on this topic: The Monetization Operation Model. Get early access 🚀 Headlines & Launches OpenAI hits $12 billion in annualized revenue (2 minute read) OpenAI has reached $12 billion in annualized revenue, roughly doubling its revenue in the first seven months of the year. The figure implies that OpenAI is generating $1 billion a month. The company has around 700 million weekly active users. OpenAI has increased its cash burn projection to roughly $8 billion in 2025, up $1 billion from its projection earlier in the year. Meta's Vision for Personal Superintelligence (3 minute read) Meta is betting on personal superintelligence as the future of computing: AI that knows you deeply, lives on devices like smart glasses, and helps you pursue your goals, relationships, and creativity. Unlike centralized visions of automation and universal basic output, Meta argues that progress comes from individual agency and empowerment. The company plans to invest heavily in infrastructure and safety to bring this technology to billions. Amazon Invests in 'Netflix of AI' Start-Up Fable, Which Launches Showrunner: A Tool for User-Directed TV Shows (7 minute read) Amazon has invested in Fable, a startup with an AI-generated TV show service called Showrunner. The money from the investment will help develop Showrunner, which allows users to create scenes (or entire episodes) of a TV show using prompts. Showrunner will be free at launch, but the company eventually plans to charge $10-$20 per month for credits. Fable's AI models have guardrails that block offensive, illegal, or copyrighted material. Mistral Codestral 25.08 and the Enterprise Coding Stack (9 minute read) Mistral unveiled Codestral 25.08, part of a full-stack coding platform optimized for enterprise needs. It addresses deployment, customization, observability, and toolchain integration gaps to enable AI-native software development across regulated and complex environments. 🧠 Deep Dives & Analysis Query Fan-Out Technique in AI Mode: New Details From Google (5 minute read) Google's query fan-out technique, available in AI Mode, Deep Search, and some AI Overview experiences, makes multiple background searches based on an initial question. It uses a large language model to interpret queries and 'fan out' to multiple related searches, including on topics the user never explicitly mentioned. The responses are then combined into a single response with links. 30% of "Humanity's Last Exam" answers are likely wrong (7 minute read) FutureHouse researchers found that nearly a third of biology and chemistry questions in the prominent PhD-level AI benchmark contain answers directly contradicted by peer-reviewed literature. The errors stem from HLE's unusual design that created "gotcha" puzzles that even confused human experts. Pricing AI Proofs-of-Concept: free pilots will kill you (6 minute read) AI agents deliver real business value, so they should be priced accordingly from day one. Companies that position their products as pilots or technical proof-of-concepts are competing on features instead of outcomes. They should charge enough to filter out the curious and co-create the ROI model that closes the deal. Customers who see how agents help save time, reduce errors, and increase throughput won't be able to argue with the business value. 🧑‍💻 Engineering & Research How Perplexity, Notion & The Farmer's Dog use AI to stay close to customers (Sponsor) You're sitting on a goldmine of feedback: tickets, surveys, reviews, but can't mine it. Manual tagging doesn't scale, and insights fall through the cracks. Enterpret's AI unifies all feedback, auto‑tags themes, and ties them to revenue/CSAT, surfacing what matters to customers. The result: faster decisions, clearer priorities, and stronger retention. 👉 See how top teams do it Agent Interaction Guidelines (AIG) (5 minute read) Agents are reshaping roles and workflows, changing how software is planned, built, reviewed, and deployed. Developer's roles have shifted to orchestrating input and reviewing output. The Agent Interaction Guidelines are a set of foundational, evolving principles and practices for designing agent interactions that integrate naturally into human workflows. They are still being written, and more is expected to be added as the community gains more experience. GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning (31 minute read) Traditional RL approaches need thousands of expensive trial-and-error runs to learn from simple success/failure scores. GEPA uses natural language reflection to diagnose what went wrong and evolve better prompts through a Pareto-based selection strategy. GEPA (Genetic-Pareto) dramatically outperforms both reinforcement learning methods like GRPO and leading prompt optimizer MIPROv2 while requiring up to 35× fewer training attempts. Latte: Decentralized Test-Time Adaptation (16 minute read) Latte is a framework for test-time adaptation in decentralized environments using pre-trained vision-language models. It combines local and shared memory to improve adaptation under limited data conditions. 🎁 Miscellaneous Something that has become undeniable this month (2 minute read) The best available open weight models new come from Chinese AI labs. The delay in the release of OpenAI's open weights model may be due to a desire to be notably better than the truly impressive lineup of Chinese models. This post lists the biggest Chinese model releases of this month, along with links to notes on each one. 'Virtual scientists' to solve complex biological problems (5 minute read) Stanford Medicine developed a team of AI agents that mimic real research labs, complete with a virtual PI, specialized scientists, and a built-in critic. The system has already designed promising COVID-19 nanobodies in days that outperformed existing antibodies in lab tests. Google Earth AI (1 minute read) Google introduced Google Earth AI, a suite of geospatial AI models and datasets designed to tackle major planetary challenges. It includes models for weather prediction, flood forecasting, wildfire detection, and urban planning. ⚡ Quick Links 28% of inbound calls go unanswered. AI can help you fix it. (Sponsor) CallRail's free guide, AI Voice Assistant 101, shows how AI answers calls anytime, captures leads, and sends conversion signals to ad platforms. 👉 Get the guide or 🚀 try it free. Microsoft in advanced talks for continued access to OpenAI tech (3 minute read) Microsoft and OpenAI are negotiating new terms that would let Microsoft continue using OpenAI's technology even after the startup declares AGI— a crucial pivot since their current contract would cut Microsoft off from advanced models at that milestone. Anthropic Backs Healthcare Interoperability (3 minute read) Anthropic joined a CMS initiative to modernize healthcare data sharing using AI. Meta's $72B Bet on AI Infrastructure (2 minute read) Meta plans to spend up to $72 billion on AI infrastructure in 2025, doubling its capex to build data centers and superclusters like Prometheus and Hyperion. Love TLDR? Tell your friends and get rewards! Share your referral link below with friends to get free TLDR swag! https://refer.tldr.tech/0b6a6dc1/2 Track your referrals here. Want to advertise in TLDR? 📰 If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us. Want to work at TLDR? 💼 Apply here or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan, Ali Aminian, Jacob Turner & Sahil Khoja Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR AI isn't for you, please unsubscribe.