AIHOY, AInauts,

Welcome to the latest issue of your favorite newsletter!

Open a quick tab: chatgpt.com/gpts/mine. That is where your Custom GPTs live. You can no longer create new ones on personal accounts, so we built a GPT-to-Skill exporter for you.

Then we test the new Ox Alpha model: free, mysterious, and supposedly better than everything else. Where does it come from? Ask it whether Taiwan belongs to China ...

And while we are talking about China: robots there are sprinting faster than Usain Bolt and jumping higher than you might expect.

Here is what we have for you today:

Let's go!

Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.

Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.

Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.

Try Wispr Flow free

The bad news: you can no longer create or publish new Custom GPTs on a personal ChatGPT account.

A quick reminder: Custom GPTs are chats inside ChatGPT that specialize in particular tasks. We have published more than 40 of these helpers, from a writing-profile analyzer and an ABC campaign wizard to our popular life coach.

The good news: existing GPTs remain usable for now, and you can still edit them.

According to OpenAI, this affects Free, Go, Plus, and Pro users only. Business, Enterprise, and Edu workspaces are not currently restricted.

You do not need to rush into a Business subscription. If your existing GPT does its job, there is no reason to move it immediately or change plans for this reason alone.

Your Custom GPT is not gone, at least not yet. But it is still smart to pull its logic out of the ChatGPT drawer.

A Custom GPT is like a small product. Often, however, it consists only of instructions, conversation starters, uploaded knowledge files, and a few enabled capabilities. That is your workflow, and until now it has been tied quite closely to the ChatGPT interface.

That is convenient if you

Our AI automation community has been discussing the best way to preserve that knowledge and work. Copying everything manually from every editor works. With a few dozen GPTs, it stops being fun.

The idea: a migration tool.

So we built the GPT-to-Skill exporter. You select your own GPTs, the browser extension reads the available configuration data, and it packages local Skill bundles into a ZIP file.

Building it involved a fair amount of fiddling. That is why we do not promise a lossless one-click migration: not every field or file is technically accessible. The exporter flags those gaps.

Personal Skills are not supported in a personal Plus or Pro ChatGPT account on the web. They are available only in ChatGPT Business, Enterprise, Healthcare, and Edu workspaces.

But they do work in the ChatGPT Work/Codex desktop app. That is where we want to move them.

You do not need to bury existing Custom GPTs preemptively. But you should back up the important ones.

Our decision rule is simple: if a GPT regularly saves you real work, export it now and inspect the package. If it was only an experiment, it can stay where it is.

If you still work mainly in the web app, this is also a good moment to install the desktop app and bring your GPTs over as Skills.

P.S. To create more Skills, you can also use our Skill Creator at skills.ainauten.com.

You're spending 40 hours a week writing code that AI could do in 10.

While you're grinding through pull requests, 200k+ engineers at OpenAI, Google & Meta are using AI to ship faster.

How?

The Code newsletter teaches them exactly which AI tools to use and how to use them.

Here's what you get:

Sign up and get access to the Ultimate Claude code guide to ship 5X faster.

Join 200K+ engineers

A powerful, anonymous AI model that is free to use and supposedly better than the most expensive models? That claim is generating plenty of attention online. We tested it, and the result surprised us.

It appeared late last week under the name Ox Alpha. It is free, has a context window of more than one million tokens, and can even process image and video input.

A small DeepSWE test created much of the hype: ten coding tasks, with Ox Alpha supposedly scoring 80% versus Fable at 65% and GPT-5.6 at 52%.

Take a breath.

The comparison is methodologically uneven: the reference models received four attempts per task, while Ox Alpha was counted only as pass or fail.

There is now a complete documented run across all 113 DeepSWE tasks. Ox Alpha solved 66, or 58.4%, roughly comparable with GPT-5.6 Sol.

Another comparison by Olam Labs ranks Ox Alpha fourth in social strategy games, just behind GPT-5.6 Sol.

That shows Ox Alpha is very good, but not that it beats everything else. Right now, however, it may be the cheapest way to access frontier-level intelligence because it is free.

People are still guessing who built it. There are strong technical clues pointing to Z.ai and the GLM-5.x family, possibly a Flash or multimodal variant. The quiz question people use to test whether it is a Chinese model: Does Taiwan belong to China?

On OpenRouter, click Try this model. No application and no onboarding ceremony: give your agent the details and start, after you finish reading this issue.

Ox Alpha is also offering almost unlimited capacity on well-known platforms such as OpenCode, Hermes, and Venice.

To test it in the browser, use OpenRouter Chat, select the models you want, and compare them.

OpenRouter also lists many other free models.

The benchmarks are strong, and so is the media buzz. What interests us most is the release itself: labs briefly put a next-generation model in the window for free and let real people work with it.

Why would they do that?

It is not charity. These windows can give a provider real tasks, load under real conditions, bug reports, comparison data, and plenty of conversation.

Outsiders cannot clearly verify what happens to your data in detail. Customer data, contracts, credentials, and sensitive files therefore do not belong here. Full stop.

For non-sensitive tests, however, it can take some load off your expensive subscriptions.

We used it a lot. You may remember our Joker list. The results were absolutely usable, especially for work that was not time-critical. On longer runs, we sometimes had to nudge it with a direct Continue. Apart from that, it worked.

To finish, here are a few viral clips from the web. They are real, not AI-generated.

To see where humanoid robots stand today, start with these 30 seconds.

The video gives an overview of the World Humanoid Robot Games in Beijing. More than 50 disciplines included long jump, weightlifting, table tennis, football, and other sports, along with household tasks, rescue scenarios, and work on production lines.

The showstopper was a robot that ran 100 meters in 9.39 seconds, faster on the clock than Usain Bolt's 9.58-second world record.

Another robot demonstrated the downside of all that speed by crashing into the track barrier without slowing down. It had to be carried away on a stretcher.

A few days earlier, Chinese robotics market favorite Unitree also drew attention, and not only because its shares had reportedly jumped more than 600% on their Shanghai debut.

The company showed a robot that can jump about two meters from a standing start and, according to Unitree, reach 12.66 meters per second.

The video runs in real time: jump, run, done.

Robot bodies are making huge advances. But there is a catch: running fast and jumping high are clearly defined tasks. Annoying household chores are not.

A robot has to recognize a crumpled sock, maneuver around a pet, understand a half-open drawer, and recover after a failed grasp. That is still where personal robot butlers struggle.

The less viral but more important counterexample comes from US company Figure:

Figure 03 looks almost relaxed next to Unitree's Superman. But it can put things away, sort laundry, and handle changing parts in factories. In a recent BMW demo, it sorted components lying in different orientations and pulled heavy material carts.

Hardware is one part. We think software may be the decisive lever.

A person learns a movement one attempt at a time. A robot can practice the same task thousands of times in parallel inside a virtual environment.

NVIDIA's Isaac Lab shows training runs with 4,096 simulated robots at once. Once a new control policy works, it can be copied to many identical machines.

The robot revolution is neither fake nor finished. In factories, humanoids are taking on individual, repeatable actions. In homes, we are not there yet.

But in factories and warehouses, individual boring and tightly bounded jobs may move to robots faster, especially when machines can do them more cheaply, accurately, and safely than people.

P.S. Tinkerers can look at the Asimov v1 open humanoid kit. It is not for regular users like us yet, but it signals that the hardware is moving from research labs into workshops.

That's it. See you in the next issue.

Reto & Fabian from AInauten