Claude Sonnet 4.5 4️⃣, Claude Agent SDK 💻, Agentic Commerce Protocol 🛒

Anthropic's latest model boasts the highest score on the SWE-bench Verified, alongside major improvements in computer use, reasoning, and math ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR Together With TLDR AI 2025-09-30 How to use contextual AI to surface relevant information, mid-conversation (Sponsor) Employees switch between apps +1,200 times per day, killing productivity. This "toggle tax" is avoidable: AI + integrations can give users the information they need, without rummaging through 37 open tabs. On October 9, technical experts from Glean and Zoom will demonstrate how contextual AI can surface relevant answers—both from your enterprise data and from the web—directly inside your Zoom conversations (or wherever employees communicate). The session covers real scenarios for Sales, Support, and Engineering teams, and how the integration keeps sensitive data private while making discovery effortless. You'll also get a rollout playbook for driving adoption and measuring ROI. Register for live + on-demand access 🚀 Headlines & Launches Claude Sonnet 4.5 (5 minute read) Anthropic's latest model boasts the highest score on the SWE-bench Verified (77.2%), alongside major improvements in computer use, reasoning, and math. The release includes major product updates: checkpoints in Claude Code for instant rollback, the Claude for Chrome extension for Max users, and the Claude Agent SDK that makes the infrastructure powering Claude Code available to all developers. Buy in ChatGPT with Instant Checkout (4 minute read) ChatGPT now supports direct purchases from Etsy sellers via Instant Checkout, with Shopify coming soon. The system runs on the open Agentic Commerce Protocol, co-developed with Stripe, and is open for merchant integrations. Anthropic launches Claude Agent SDK for building versatile AI agents (11 minute read) Anthropic introduced the Claude Agent SDK, expanding its capabilities beyond coding to enable building versatile agents for tasks such as finance management and customer support. This SDK equips agents with tools for context gathering, executing tasks, and iterating based on feedback, optimizing workflows through features like bash scripting and subagents. Developers can leverage the SDK for immediate integration and enhanced agent functionality. 🧠 Deep Dives & Analysis How GPU Matmul Kernels Work (96 minute read) The architecture, assembly, and kernel design techniques behind high-performance matrix multiplication on NVIDIA GPUs. A Research Agenda for the Economics of Transformative AI (32 minute read) There are nine "Grand Challenges" for studying the economic effects of AI, spanning issues from income distribution to AI safety to human meaning. Economists need to start researching these issues now due to the speed of AI progress and the unprecedented resource flow towards the AI industry. By the time transformative AI, which they define as driving 3x productivity growth, arrives, it will be too late for economists to aid in the upheaval. Building AI for cyber defenders (9 minute read) AI models are now useful for cybersecurity tasks in practice, not just theory. Anthropic has caught hackers using their tools and invested in making Claude Sonnet 4.5 useful for defenders looking to detect, analyze, and remediate vulnerabilities in code and deployed systems. Now, the defenders need to adopt and experiment with AI to keep pace. 🧑‍💻 Engineering & Research The SANS Blueprint for Secure AI (Sponsor) The SANS Secure AI Blueprint provides a proven model for reducing risk and owning AI securely. Structured around three imperatives — Protect AI, Utilize AI, Govern AI — it helps SOCs, CISOs, and engineers defend GenAI in real-world deployments. Get the blueprint or explore additional AI security resources from SANS. Agentic Commerce (Website) This site features design flows for embedded commerce in ChatGPT. OpenAI's Agentic Commerce Protocol is an open standard that enables purchases. It enables agents to reason over structured state, invoke tools at each step, and keep customers informed in real time. This page will help developers start with the essentials, deepen their understanding, and prepare for production with focused ACP resources. DeepSeek-V3.2-Exp: Sparse Attention Model (4 minute read) DeepSeek-V3.2-Exp has introduced a sparse attention mechanism aimed at improving training and inference efficiency on long-context sequences. LoRA Without Regret - Thinking Machines Lab (27 minute read) LoRA, the most popular method for fine-tuning AI models, can match the performance of expensive full fine-tuning when applied correctly, delivering the same results at two-thirds the cost. The key is applying LoRA to all network layers (especially MLPs rather than just attention layers) and ensuring the rank is high enough for the dataset. Thinking Machines found ranks of 256-512 work well for typical post-training scenarios combined with 10× higher learning rates than fine-tuning. 🎁 Miscellaneous OpenAI's New Sora Video Generator to Require Copyright Holders to Opt Out (5 minute read) OpenAI's new Sora video generator can create videos that feature copyright material unless copyright holders opt out of having their work appear. The company began alerting talent agencies and studios about the product and its opt-out process over the past week. The model will be released in the coming days. It won't generate images of recognizable public figures without their permission. How is it possible that Claude Sonnet 4.5 is able to work for 30 hours to build an app like Slack?! (3 minute read) Claude Sonnet 4.5's system prompt reveals how it is able to work for 30 hours. This post goes into detail about how it works. A copy of the system prompt is available in the thread. ⚡ Quick Links Not actively job hunting? Great, this a16z backed startup does it for you (Sponsor) Dex is your AI recruiter for software engineers, finding you $200k-1m tech jobs in just 15 minutes. Chat to Dex and he'll scan thousands of roles, connecting you with hiring managers, and helping you negotiate the compensation you deserve. No more job boards, no wasting time speaking to endless recruiters. Don't wait—talk to Dex today, totally free! Do Humans Really Have World Models? (4 minute read) There's little difference between the way humans and large language models think. HunyuanImage-3.0 Released (GitHub Repo) Tencent has open-sourced HunyuanImage-3.0, providing both inference code and model weights alongside a detailed technical report. Vibe Check: Claude Sonnet 4.5 (2 minute read) Claude Sonnet 4.5 is faster than GPT-5 Codex, and smarter and more steerable than Opus 4.1. Logics-Parsing (GitHub Repo) Logics-Parsing is an end-to-end document parsing model built on a general Vision-Language Model that excels at accurately analyzing and structuring highly complex documents. 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 Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR AI isn't for you, please unsubscribe.