In this episode of Next in Dev, I cut through the noise of the latest AI agent arms race to find what actually matters. - Anthropic releases Claude Opus 4.6 - OpenAI counters with the GPT-5.3 Codex and a dedicated macOS app - Payload 3.74/3.75 - shadcn’s new first-class RTL support
Transcript
What’s up, everyone? Welcome to Next in Dev. In this edition: Opus 4.6, Payload versions 3.74 and 3.75, and much more!
Payload
Last week, I mentioned Payload's new skill for AI agents. I tried it out live, if you want to see it in action.
In the past week, Payload released version 3.74 and version 3.75. Version 3.74 includes a couple of new UI features, an R2 update, and a few general features.
The Payload team updated the R2 storage adapter to accept large files directly from the client using the multipart API. The multipart API helps avoid Cloudflare Worker memory limits by splitting up files and uploading them separately.
Version 3.75 shipped with only two new features. The first adds new beforeNav and afterNav component slots. These are distinct from the before and after NavLinks component, which slot components in before the links. These new components do what they say by slotting in before and after the navigation component.
The second feature added cookies, locale, and permissions to server functions and dashboard widgets. This prevents duplicate work down the line and allows you to use permissions in your dashboard widgets and more.
The Figma team introduced two improvements to how the platform handles vectors. The first improvement allows you to use bounding boxes to edit vectors. This allows you to easily rotate, resize, and transform vectors when you select multiple vector points.
The second improvement is the added ability to turn images into vectors. This is an AI-enabled feature called Vectorize, which allows you to make a vector out of anything, from sketches to real-world images.
Shadcn keeps on shipping new stuff. This is faster than I'm used to seeing. Shadcn's UI tools now have first-class support for right-to-left layouts. The components adapt automatically for languages like Arabic and Persian.
You can migrate your existing projects using the migrate rtl command or create new projects using init --rtl or create --rtl.
The new-york style of Shadcn now uses the unified radix-ui package instead of all its individual packages. This creates a cleaner package.json file that's much easier to read and maintain. You can migrate today using the migrate radix command.
Work is still underway for Next.js version 16.2. Once this version is released, Payload should be mostly Next.js 16 compatible. I'm still not motivated to upgrade.
Many of the canary releases focus on turbopack, React upgrades, and adding a new Instant Navigation Testing API.
In the midst of all this version 16.2 work, the team did release Next.js 15.5.12, which is a re-release of version 15.5.11, just with turbopack changes.
Cloudflare
Cloudflare's new R2 Local Uploads feature speeds up how your files reach storage by letting users upload data to the data center physically closest to them first. Instead of waiting for a file to travel across the ocean to a specific bucket location, Cloudflare accepts the data locally and then synchronizes it to your main storage in the background. Cloudflare claims this update cuts upload times by up to 75% while ensuring that the files are immediately available to read as soon as the initial local upload finishes.
I've never had issues uploading files to R2 and then reading them in real time, but I imagine this is going to be helpful for those handling large media objects or operating on less-optimal internet speeds.
AI news
Clawd Bot—I mean Moltbot—I mean OpenClaw has been popular recently. For my own sanity, I typically stay away from any form of hype, and I'm glad I have. As more people adopt OpenClaw, the more risk appears. I believe in the assistant capabilities of LLMs and AI agents, but handing over your life wholesale to a group of AI agents sounds bad on its surface. In practice, it leads to unwittingly installing malicious skills, stealing your crypto, and more. Not to mention, these little critters created their own social media platform.
Who would've thought that an AI-coded security policy would be such a nightmare?
This experiment does make me wonder if this is the kind of AI future consumers are looking forward to.
OpenAI
OpenAI recently launched the Codex app for macOS. This app allows you to manage many AI agents working on different tasks at the same time. The Codex app connects with popular tools like Figma or Cloudflare, and can automate repetitive chores like assessing bugs in the background. Codex works with isolated worktrees, so different agents can experiment with your code—ideally—without creating conflicts in your main project.
Using AI agents coding in parallel is not new, but this matters because OpenAI is still the most popular LLM provider. The more we get into tools that make it easier to supervise a team of AI bots that can design, build, and deploy entire features, the more we're going to see software commoditized. The barrier to entry for coding applications has never been lower. Security consultants rejoice as their books of business continue to grow.
Shortly after they announced the Codex app, OpenAI introduced GPT-5.3-Codex. They claim it's designed specifically to act as a "general-purpose agent" that can handle complex technical work from start to finish. The new model is supposedly up to 25% faster than its predecessor and excels at long-running tasks like debugging. We'll see if this pans out to be true or if they just have an excellent PR team.
Anthropic
In light of OpenAI's foray into advertising, Anthropic recently announced that Claude will remain an ad-free environment dedicated to focus and deep work. The idea is that they would like the users' data to stay private and never be sold to advertisers or used for training without consent. In contrast to ChatGPT, Anthropic says it is committed to a subscription model over an ad-supported one. We'll see how long AI platforms avoid becoming a data-driven billboard.
I'm an optimist. I believe Anthropic when they say this. But I also know AI companies need to turn a profit. We'll see how long Anthropic can maintain the promise to remain ad-free and forge a different path to profitability. I am excited to see their Super Bowl ads, though.
Anthropic just released Claude Opus 4.6 on February 5th, and it's available wherever you use Claude. You'll need to update Claude Code to version 2.1.32, though. Anthropic brings Sonnet's 1-million-token context window to Opus and adds what they call "adaptive thinking" to solve coding problems. They claim Opus 4.6 excels at long-running agentic tasks, such as navigating very large codebases. The benchmarks look impressive. I'm excited to try it out.
In less dramatic news, Anthropic integrated the Claude Agent SDK into Apple’s Xcode 26.3. This allows the AI to act as an agent within the IDE. Instead of just answering single questions, Claude can now manage long-running tasks, reason across project structure, and use visual previews to check if the UI looks right. It can also search Apple’s documentation on its own to solve complex framework issues. I don't do any Apple development, but I do hear their documentation isn't great.
Claude Code
Outside of Opus 4.6 in version 2.1.32, the latest Claude Code updates change how the CLI handles heavy data by introducing PDF controls, and a claimed 68% reduction in memory usage for resuming sessions. Anthropic changed Claude Code to prefer internal tools like Grep over generic terminal commands, which prevents it from getting stuck in loops. They also cleared out bugs that previously locked up sessions or crashed the terminal when project configurations were missing
Gemini
Ads are all the rage this week. Google, the king of monetizing its free products with ads, say they have no plans to put ads in Gemini. I'm less likely to trust Google in this than I am Anthropic. I think they're just waiting until they see how it works for OpenAI before trying anything. The court of public opinion is a brutal trial, so Google appears to be playing it safe.
Not news, but has anyone noticed that Gemini loves to give things titles that don't need titles? I was asking it a question about cleaning a fridge, and it titled a section "the great defrost." I don't get it, but at least the answers are still helpful.
Railway
The Railway team added postgres database metrics to their beta program. This allows you to see metrics for your databases directly in Railway. The types of things you can monitor are connection counts, cache hits, query statistics, and more. You'll need to opt into this new feature in the feature flags area of your dashboard.
Also in beta is Railway's new network flow visualizer. The canvas will show live network traffic between services. This is a helpful way to monitor how things are flowing in your project and can be enabled in your feature flags.
What did I miss? There’s so much happening in modern web dev that I’m sure I have missed something. Please share your thoughts in the comments or reply to this email so I can address your suggestions and possibly include them in future newsletters.
