Welcome back. If you’ve been thinking about using OpenAI’s new Sora video app, you might want to hold that thought. If at some point you decide to delete your account, some users are sharing screenshots that show you’ll be deleted and banned from virtually all OpenAI products — including ChatGPT and API usage.

Today: Apple unlocks in-app AI features for builders, ChatGPT gives developers access to its 700M users with payments API, and Anthropic’s new SDK for agents.

Welcome to The Code. This is a 2x weekly email that cuts through the noise to help devs, engineers, and technical leaders find high-signal news, releases, and resources in 5 minutes or less. You can sign up or share this email here.

ChatGPT now lets agents purchase on behalf of users with payments API: The open-source Agentic Commerce Protocol, co-developed with Stripe, defines REST endpoints and webhooks that merchants can implement to accept AI-initiated transactions while maintaining control over fraud detection and customer relationships. Developers can integrate with a single API to reach 700 million weekly ChatGPT users. Stripe customers need just one line of code to enable the feature.

Anthropic drops Claude 4.5 Sonnet and Agent SDK: The new model claimed top spot on the SWE-bench coding benchmark with 77.2% accuracy. The model can maintain focus for over 30 hours on complex multi-step tasks, with early adopters like Canva seeing dramatic improvements in their engineering workflows. Pricing comes at the same $3/$15 per million tokens as its predecessor. Developers are also getting access to the Claude Agent SDK (the same infrastructure powering Claude Code) to build their own autonomous coding agents.

Apple ships free on-device LLM access with just 3 lines of Swift code: Apple just opened its 3-billion parameter language model to all developers through the Foundation Models framework in iOS 26. The framework's killer feature: Native Swift integration using the @Generable macro for type-safe outputs—no JSON parsing needed. Developers get Apple Intelligence's core model with zero API costs and complete offline functionality. The framework includes guided generation, tool calling, and streaming APIs.

Real AI Agents and Real Work: OpenAI's latest benchmark shows AI models matching human experts on real 4-7 hour professional tasks across finance, law, and retail. Wharton professor Ethan Mollick demonstrated this leap by getting Claude to replicate an entire economics paper in minutes, complete with statistical analysis and data conversion.

The secret to making AI coding agents actually useful? Let them run wild: Developer Simon Willison just coined an idea that's reshaping how we think about AI coding tools — "designing agentic loops." The core insight: modern coding agents like Claude Code and Codex CLI work best when you let them fail fast and iterate freely in what Willison calls "YOLO mode" (yes, really).

A theory of the AI market: RunLLM's founders just published their year-old AI market predictions with a twist — they were mostly right. Their core thesis: enterprises are about to shift from AI experimentation to demanding measurable ROI, creating a barbell distribution of value in the market.

Quiet Influence — A Guide to Nemawashi in Engineering: Technical proposals often fail not due to poor engineering but because of weak social strategy. Staff engineers succeed by using "Nemawashi," a technique that builds consensus before formal presentations. Instead of surprising stakeholders with a complete proposal, meet with key people individually first.

Anthropic reveals the secret sauce behind AI agents that actually work: Forget prompt engineering — context engineering is the new game in town. Anthropic's engineering team just dropped a masterclass on why AI agents fail at complex tasks (and how to fix them).

How to allow an agent to dynamically self-improve its own memory using Sonet 4.5: Anthropic’s new model treats memory like local files that the model can manipulate through tool calls. The system lets Claude create memory directories, edit specific lines, and even delete outdated information — all stored client-side on your machine rather than Anthropic's servers.

9 principles for building modern developer documentation in the AI era: In this comprehensive guide, the author emphasizes that documentation must be fast (static pages, optimized assets), readable (concise, example-heavy, skimmable), and helpful (includes workarounds, migration guides, automated link checking).

LoRA Without Regret: Thinking Machines researchers demonstrate that LoRA (Low-Rank Adaptation) can match full fine-tuning performance under specific conditions. LoRA achieves equivalent results when applied to all network layers and given sufficient capacity relative to dataset size, enabling efficient post-training with 10x higher optimal learning rates than full fine-tuning.

DeepSeek-V3.2-Exp: DeepSeek researchers introduce DeepSeek Sparse Attention, using a lightning indexer for fine-grained token selection in long contexts. The model achieves 2-3x speedup on 128K sequences while maintaining performance parity with DeepSeek-V3.1-Terminus.

Pretraining LLMs with NVFP4: NVIDIA researchers demonstrate stable 4-bit precision training of large language models using NVFP4 format with novel techniques. A 12B-parameter model trained on 10 trillion tokens achieves FP8-comparable accuracy, establishing the first public evidence of sustained 4-bit pretraining at multi-trillion-token scale.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team