Welcome back. In today's agent-first world, voice AI remains the weakest link. It can handle simple commands, but tends to fall apart the second it needs to think through a complex problem. OpenAI just shipped new voice models in the API to bridge that gap. And it wasn't the only thing they launched yesterday.

Also: Senior developer's guide to slash API costs with Codex, build a self-improving knowledge base in Obsidian, and find the contradictions every AI jobs report is hiding.

You are the reason our team spends hundreds of hours every week writing this email. This newsletter exists to serve you! Please answer a few short questions below to help us create better emails for you:

→ Take our 60-second reader survey

OpenAI brings real-time reasoning to voice agents: The ChatGPT maker just dropped three new audio APIs, featuring GPT-Realtime-2, which offers GPT-5-level reasoning during live speech. It can handle multiple tools simultaneously, “think” while speaking, and outperforms its predecessor by 15 points. Along with a 70-language translator and a streaming transcription model, OpenAI’s Codex is now available as a Chrome extension for macOS and Windows, running across tabs in the background. Try it here.

Cursor unveils end-to-end PR reviews inside the editor: The AI coding startup just shipped version 3.3, letting teams manage pull requests from creation to merge without a context switch. New Reviews, Commits, and Changes tabs centralize inline comments and history, while "Build in Parallel" uses async subagents to execute plan steps simultaneously. Plus, a new shortcut automatically breaks down massive diffs into smaller, mergeable PRs.

Prime Intellect debuts an open stack for self-improving agents: The San Francisco-based AI startup just released Lab out of beta, providing an all-in-one platform for reinforcement learning. You can now build environments, run evals, train adapters, and deploy in one place. Teams can fine-tune 14 models from Nvidia, OpenAI, Meta, and Qwen with pay-as-you-go pricing and zero GPU management. During the beta, users completed over 10,000 jobs in automation, data, and coding.

Blitzy is the only autonomous software development platform built for enterprise codebases, enabling you to accelerate the SDLC by 5x without sacrificing quality or compliance.

Co-pilots can’t operate at the scale Blitzy does:

Start building today.

The labs are racing. The bet behind “world models” is that AI learns best by observing and interacting with the physical world. In March, Yann LeCun raised $1.03B for AMI Labs on this exact premise. The "godmother of AI", Fei-Fei Li, and other top rivals are chasing the same goal with their own labs.

Pixels are the textbook. While an LLM predicts the next word, a world model predicts what happens next after an action. By pairing dashcam footage with the actual steering inputs, the model learns exactly how every turn of the wheel affects the road. This allows a self-driving system to log thousands of hours of practice in a simulation before the car even hits the pavement.

Data is the constraint. World models need a rare combination of video data paired with the specific actions that generated it. Driving and gaming have these labels built in, but everything else requires manual recording on physical hardware. VC James Wang calls this "data friction," and it’s the biggest bottleneck keeping this tech from scaling past a few narrow use cases.

Where the bet pays off. World models give AI a practical understanding of physics and consequences that language alone can't replicate. If your team is building anything for the physical world, this is the architecture to watch in 2026. You can find the full technical case in LeCun's paper.

Codex used to stall every few minutes, waiting for permission to run scripts or access the network. These constant interruptions made background sessions nearly impossible because the agent would sit idle until someone clicked approve.

OpenAI's head of Codex just shared a fix. Auto-review is now the default inside OpenAI and delegates each approval to a separate agent that vets the action against a risk policy before letting it run. The shift reportedly cuts user prompts by roughly 200x. To enable it, add this to your “~/.codex/config.toml”:

Restart Codex, and the reviewer takes over. The TUI shows each decision in real time, so you can still see what the reviewer approved or rejected.

P.S. You can find 50+ AI coding hacks here.

How to run LLMs locally and in the cloud: This tutorial shows you how to run open-source LLMs like Gemma and GLM locally or in the cloud. You'll learn how to plug these models into different coding setups, compare how they stack up for dev work, and build out your own environments for AI-assisted coding.

Raindrop: The monitoring platform for AI agents. Just as Sentry tracks errors in your web apps, Raindrop catches silent agent failures in production.

OpenAgents by Vercel (5.2k ⭐): This repo is an open-source template for building cloud agents. It offers a complete toolkit to turn prompts into code changes. This includes a web UI, agent runtime, sandbox orchestration, and GitHub integration, all running in the cloud, so you don't have to keep your laptop on.

Toward a native foundation model for multimodal agents: AI agents often struggle because they treat vision as an afterthought rather than a native skill. To solve this, researchers developed a framework that processes video and webpages naturally. By combining better visual perception with multitask reinforcement learning, they've improved performance in coding and complex GUI tasks.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 270K+ engineers and 150K+ followers on socials. Get in touch.

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

Until next time — The Code team