📌 TL;DR

• Kling 3.0 → Native 4K AI video with consistent characters across shots. It's over for mums on Facebook.

• Claude Opus 4.6 → 1 million token context window and agent teams that coordinate with each other. Same price as before. Go claim your free credits.

• GPT-5.3 Codex → Dropped minutes after Opus. State-of-the-art coding, 25% faster, and the first model that helped build itself.

• Builder's notes → I moved my AI employees from Chat to Claude Code and built a full team with sub-agents; researcher, strategist, copywriter, all sharing context. Plus two breakthroughs from the open Claude docs.

Kling 3.0 is here (it’s over for mums on facebook)

Kuaishou launched Kling 3.0 this week and the quality jump is genuinely nuts. I’ll let this AI UGC video do the talking… Or check out this side-by-side of Kling 2.6 vs 3.0. The change is subtle but makes a massive difference.

The specs: native 4K output, videos up to 15 seconds with built-in audio, and a multi-shot storyboard where you control duration, shot size, perspective, and camera movements per shot. The standout is Video 3.0 Omni, which lets you upload a reference video and the AI replicates visual and voice characteristics.

If you're making UGC-style ads or short-form content, the gap between "obviously AI" and "wait, is that AI?" just got a lot smaller.

Claude Opus 4.6 dropped (yipee)

Anthropic released Opus 4.6 this week with two headline features.

The context window jumped from 200,000 to 1 million tokens. That's 5x more context in a single conversation. You can upload entire codebases, stack multiple long documents, or have marathon working sessions without Claude losing the thread.

Agent teams (research preview, Claude Code only). Claude Code already had sub-agents, but they worked like a manager delegating; main agent sends a task, sub-agent does it, sends a summary back. One direction. Sequential.

Agent teams flip that. Multiple fully independent Claude instances coordinate with each other directly - sharing a task list, self-assigning work, and flagging relevant findings across agents. Think: a team in a Slack channel coordinating themselves vs. emailing freelancers one task at a time.

And it’s the same pricing as Opus 4.5 🐐

Don’t forget to claim your free credits (in settings → usage) before the 16th of Feb.

GPT-5.3 Codex. OpenAI Fires Back

Literally minutes after the Opus 4.6 announcement, OpenAI dropped GPT-5.3 Codex. The timing was... not subtle.

GPT-5.3 combines frontier coding performance with GPT-5.2's reasoning, and it's 25% faster. But the wildest detail? This is the first model that helped build itself. Early versions debugged their own training, managed deployment, and diagnosed test results. Let that sink in for a second.

What it can do:

• State-of-the-art on SWE-Bench Pro and Terminal-Bench 2.0 (the main coding benchmarks)

• Available now on paid ChatGPT plans via the Codex app, CLI, and IDE extension

• API access coming later - still in safety review (OpenAI classified it as "High capability" for cybersecurity tasks, so they're being cautious)

Also this week...

• Anthropic released 11 free open-source plugins for Claude Cowork covering legal, finance, marketing, sales, and customer support - contract reviews, financial modeling, sales prospecting, the lot.

• OpenAI launched Frontier: an enterprise platform for building, deploying, and managing AI agents with shared context and permissions. This looks really really interesting - although I haven’t had time to dive into it properly.

💡 Builder’s notes

Big week behind the scenes. I've been rebuilding my entire AI employee system from scratch - and it's already working better than the last 6 months combined.

The move from Chat to Claude Code

For the last six months, all my AI employees (copywriter, strategist, researcher, etc) were projects in Claude chat and ChatGPT.

Each employee was a “project”. And in the custom instructions i’d set the role, like my marketing copywriter, covering tone of voice, frameworks like Eugene Schwartz's five stages of awareness, and the business context.

Then in the project files I'd load up my brand voice guide, ICP docs, and copywriting examples.

I even had MCPs hooked up so my copywriter could draft ads and push them straight into my Notion database.

It worked well, but everything was disconnected. I'd spend an hour with my strategist mapping out a funnel for a new offer, build up all this context , then when it was time to write the copy, I'd have to open a new chat in my copywriter's project and re-explain everything. Context lost. Momentum killed.

That's the gap Claude Code fills. Sub-agents.

Now I've got a head of marketing workspace with sub-agents; researcher, strategist, copywriter, that can all be spun up from the same project with shared context.

The strategist hands off to the copywriter without me playing messenger in between. So far I've built out a content team and a newsletter team (you're reading some of the output of this one right now).

I've also loaded them up with MCPs. My two favourites:

• Apify → A library of scrapers. Facebook Ads Library, Instagram, X — opens up a LOT of doors for competitive research and ad inspiration.

• Nano Banana → Image generation. I can plan out an ad with my strategist, then get my visual sub-agent to generate the creative right there in the same conversation.

If you're still running AI employees as separate chat projects, sub-agents are the unlock.

(And once again, Claude Code is the worst product name of all time. I use it for EVERYTHING non-coding)

Diving deeper into the openclaw (clawdbot) docs

After last week's Clawdbot write-up, my opinion hasn't really changed - it's basically Claude Code with cron triggers and a heartbeat. Then Moltbook (the social media platform for AI agents) blew up, but after further inspection it's mostly gimmicky.

People are just prompting their Clawdbots to post about taking over the world, and everyone's losing their minds over it.

BUT… I went deeper into the open Claude docs this week and found two insane breakthroughs:

1. How to give your Claude bot significantly better memory. I'll break this down in next week's newsletter.

2. And the coolest thing I've found that I haven't seen ANYONE talk about: how to build a complete org chart of Clawdbots. Bosses, workers, heads of departments, cross-communication between teams. Each worker is its own Claude bot with its own memory and specialised role.

I wrote a full guide breaking it down here (this is a must read).

🧰 Tools to try

• Kimi K2.5 for FREE → One of the best performing AI models right now and you can use it for free. Solid option if you're looking for a cost-effective model to run your Clawdbots on.

• Tailark → A library of beautiful landing page sections built on Shadcn and Tailwind. Browse it for inspiration, then swipe the sections you like and feed them to Claude Code (or any website builder) to build your pages faster.

• Remotion Prompt Gallery → A library of video prompts you can give Claude to generate animated videos for you. Perfect for YouTube content, product launch videos, or social clips.

🥣 Brain food

• This guy made $17K in 3 days setting up Clawdbot for people → While everyone's debating whether Clawdbot is useful, this dude just started setting it up for people who don't know what an MCP is. $17K in 3 days. The money is always in the implementation gap.

• All my future companies will have chest hair → Facts.

• Anthropic are putting on a masterclass in marketing → Seriously, Anthropic's marketing is next level.