I predicted Claude would cave

hello friends 👋,

Last Sunday I wrote that Anthropic would extend Fable 5 again.

My exact words were:

“They almost have to.”

Less than a week later, they did.

But Anthropic was not the only company feeling the pressure.

China just released a model that competes with the best models in the US, OpenAI now has its own version of Cowork, and Claude artifacts are slowly becoming real apps.

Competition is finally doing what it is supposed to do: giving us better models for less money.

Fable 5 is staying.

Starting July 20, it will be included with Max and Team Premium plans at up to 50% of their limits. Pro and Team Standard users get a one-time $100 credit instead.

Kimi K3 is the real deal.

It ranks #1 for frontend coding and #4 overall on Artificial Analysis. The main caveat: it feels sloooow.

ChatGPT Work is OpenAI’s answer to Claude Cowork.

It can research, use connected apps, work with local files, and create finished documents, spreadsheets, presentations, reports, and websites.

Claude artifacts can now use MCP connectors.

You can build an app that connects to each user’s own Slack, calendar, Notion, or internal tools without building every integration yourself.

Content OS is nearly ready.

I’m aiming to release the first version next week.

Anthropic announced that Fable 5 will become a standard part of Max and Team Premium subscriptions from July 20, using up to 50% of weekly limits.

Pro and Team Standard users will still need usage credits, but they are receiving a one-time $100 credit.

Anthropic also extended the temporary 50% increase in Claude Code limits until August 19.

So I was half right.

Fable is not becoming standard across every subscription, but Anthropic clearly realized that removing it completely was no longer an option.

GPT‑5.6 made the old pricing difficult to defend. Then Kimi K3 arrived and added even more pressure.

My cynical take?

I cannot prove the original government shutdown was a marketing stunt. The export controls were real, but the entire scarcity cycle did more for Fable’s launch than any normal campaign could have.

Release the best model.

Remove it three days later.

Bring it back for a limited time.

Extend it three times.

Then finally add it to subscriptions.

Also, Fable 5 and Mythos 5 use the same underlying model. The main difference is the level of safeguards wrapped around it.

Either way, I’m happy Fable is staying because I’ve become completely dependent on it.

Kimi K3 might be the most important release this week.

It is a 2.8-trillion-parameter model with native vision, a one-million-token context window, and strong long-running agent capabilities.

It currently ranks:

→ #1 on Frontend Code Arena
→ #4 on the Artificial Analysis Intelligence Index
→ #1 on AutomationBench-AA

It also recreated an interactive version of macOS that went viral.

Play with Kimi’s macOS here.

The pricing is where things get interesting:

So Kimi is 50% cheaper on output than GPT‑5.6 Sol and 70% cheaper than Fable 5.

There is one caveat.

It feels slow.

Interestingly, the raw generation speed is not actually much worse. Artificial Analysis measured Kimi at around 62 tokens per second and Fable at 60.7.

The problem is that Kimi is extremely verbose. It used roughly 130 million output tokens across the Intelligence Index evaluations, compared with 87 million for Fable.

So it does not necessarily type slower.

It just thinks and writes a lot more before finishing the job.

Moonshot says the full weights will be released by July 27. Until that happens, it is technically not open-weight yet.

And even when they arrive, you are not casually running this on your MacBook. Moonshot recommends infrastructure with at least 64 accelerators.

Basically, you can run it locally if your computer is the size of a room.

Still, this is the pattern I keep seeing.

Chinese and open models used to sit several months behind the best private models. That gap is shrinking incredibly fast.

Soon, frontier intelligence will not belong to the big dogs.

OpenAI also released ChatGPT Work.

The simple explanation:

Chat is for asking questions.

Work is for handing over an outcome.

You can ask it to research a market, analyze connected files, create a spreadsheet, build a presentation, produce a report, or turn an idea into a small website.

On desktop, it can also work with local files, desktop apps, and a built-in browser, all with your permission.

It can keep longer projects organized and run scheduled or recurring tasks.

So yes, this is basically OpenAI’s answer to Claude Cowork.

The interesting part is that OpenAI has placed Chat, Work, and Codex inside the same desktop app:

→ Chat for quick questions
→ Work for research and finished deliverables
→ Codex for software development

AI products are moving away from “here is a chatbot.”

They are becoming operating systems for getting work done.

This update is slightly niche, but potentially huge.

Claude artifacts can now connect to MCP servers.

In simple terms, you can build a dashboard or tool inside Claude that connects to the viewer’s own services.

For example:

→ a dashboard reading from their Notion
→ a task manager connected to Asana
→ a calendar tool that creates events
→ an internal app pulling information from Slack

Each user connects and authorizes their own account.

That means you do not need to build custom authentication and backend logic for every integration before sharing something useful.

The current limitation is that MCP access is available on paid plans and does not work inside publicly shared artifacts.

Still, artifacts are moving from cool demos to lightweight software apps.

I recently discovered CLIProxyAPI, and it has become one of my favourite tools.

The dumbed-down explanation:

It is a small local traffic router for AI models.

It lets Claude Code communicate with Claude, GPT, Kimi, Gemini, and other models through compatible API formats.

My current setup uses Fable 5 as the orchestrator and GPT‑5.6 as the executor.

Fable reads the project, makes the plan, and handles decisions that require taste.

GPT‑5.6 performs the edits, runs tests, and handles the token-heavy execution.

The outputs have been great because I’m giving each model the job it is best at.

I also made a complete tutorial showing the same setup with Kimi K3 as the executor:

Run Kimi K3 inside Claude Code →

The proxy is free and open-source.

Just remember that “local proxy” does not mean “local model.” Your requests still go to whichever AI provider is running the model, so be careful with private client code.

My favourite skill this week is Apple Design from Emil Kowalski’s design skills repo.

Instead of telling your coding agent:

“Make this website look more premium.”

You give it actual design principles distilled from Apple’s WWDC design talks.

It teaches the agent about:

→ interruptible spring animations
→ motion that inherits the user’s velocity
→ spacing and typography
→ translucent materials and visual hierarchy
→ reduced motion and accessibility
→ the tiny interaction details that stop a website feeling like AI slop

It is basically a taste layer for your coding agent.

And it has already made a noticeable difference in how my interfaces feel (been using it for Content OS).

A lot of you replied to my last email asking to test Content OS 🎉

I’m currently fine-tuning the script-writing system and competitor tracker, and I’m aiming to release the first version next week.

A few of my favourite features so far:

→ Lab: create and refine reusable AI skills

→ Video Lab: paste almost any video URL and get a free transcript and structural breakdown

→ Competitor intelligence: your competitors’ best-performing videos become evidence for your scripts—not blind copies

→ Daily briefs: three content ideas based on what is breaking out in your niche

→ Voice matching: every script evolves from your existing content and feedback

I do not want Content OS to become another LLM wrapper ☹

The goal is to build something that learns alongside you.

The more you post, the better it understands your voice.

The more competitors it tracks, the better it recognizes what is working.

The more feedback you give it, the better the system becomes.

If you want to test Content OS when it is ready, reply “CONTENT OS” to this email (if you haven’t already).

And if there is something you desperately want it to do, tell me. I’m still deciding what makes the first release.

That’s it for this week.

See you next Sunday,