ChatGPT's Best New Model Is Here. But You Can't Use It.

ChatGPT's best new model is here. You can't use it. Sorry!

And no, before everyone starts shouting, GPT-5.6 has not been banned. That's the hype version going around on social media. And as always it’s not quite right.

Instead it’s actually MORE worrying. OpenAI launched it as a limited preview at the US government's request, and normal ChatGPT users are outside the preview gates. We’re locked out.

This is the new precedent.

Gulp.

I said on the livestream this might be one of the darkest weeks in AI. I stand by that. Not because AI is over. The opposite. Because the best AI is shifting from app-store product to permissioned infrastructure. It’s getting TOO good.

Ignore the ban hype

OpenAI announced three models: Sol, Terra and Luna. Lovely celestial branding. Sun, Earth, Moon. Very tasteful…

Totally useless for us though.

Because we don’t have access. OpenAI says it plans broader access later, but for now GPT-5.6 is a limited preview for a small number of trusted partners in Codex and the API.

We don’t know the list or how to (har har) get on it.

Sam's follow-up is here too:

Annoying. BUT at least Sam is communicating and acknowledges that this is a messy situation. Dario Amodei and Anthropic are not communicating with their users.

Two weeks. Same pattern.

On June 12, Anthropic disabled Fable 5 and Mythos 5 after a US export-control directive. On June 26, OpenAI shipped GPT-5.6, but gated access before normal users could touch it.

Two weeks. Two frontier-models barred from general access.

Fable/Mythos was launched and quickly pulled. Mythos is apparently coming back for approved entities, not for individuals. GPT-5.6 launched in a restricted preview. Slightly different mechanism but same basic result. The average Joe is out of the loop. Sorry Joe.

This is no longer just "can I afford the good model?" It is: are you on the list? Reminds me a bit of certain Rolex watches and Hermes birkin bags - it doesn’t matter how rich you are. That’s not how you get access. This is not a model I’m happy to see spread!

There is a real reason

The government is not just doing this for a laugh. Or because they are idiots. And not just because Dario Amodei has been stirring up fear.

No, there are genuine very real concerns.

These models are getting scary-good at cybersecurity work, long-horizon agent tasks and scientific discovery. All good stuff.

But we also don't need to pretend this is normal software anymore.

All the same mechanisms for good work can be used for bad work. If it can do one it can do the other.

If a model can help discover vulnerabilities, assemble exploit pieces, run agents more efficiently and compress difficult technical work then governments are going to care. They should care. Banks, infrastructure, health systems and government departments are mostly held together with Excel sheets, procurement forms and vibes.

And even with “dumbed” models bad actors were having an absolute field day.

No model is immune to be jailbroken. And to believe this is possible (as the White House seem to be demanding…) is naïve.

All models will get broken open to be used for naughty deeds. That’s part and parcel of this I’m afraid. Up until now this has been a nuisance but these (much) smarter models make this a critical risk.

Build a model-access hedge

OK let’s come back to what do YOU do again. Always we need to come back to this.

This is not a "delete ChatGPT and live in a bunker" issue. (Although if you do have space in your bunker for one more I’d love to know…)

Use the best models when you can. Obviously. They are absurdly useful. But don't build a your life or business around a single-model.

Have ChatGPT as one option, not your only option. Keep Claude, Gemini or another hosted provider warm. Have a poke around with Chinese-hosted models like GLM through Z.ai if you are comfortable with the trade-offs. Start testing local Gemma-style or Chinese models for lower-risk work you want to fully control. I wrote about first steps with local AI models a few days ago.

Also build out an AI brain. Get everything in one place. Your prompts. Your specs. Your source files. Your evaluation examples. Your workflow notes. Your project context. Do not trap the brain of your business inside one rented interface.

Get it into a Github repo that you can move BETWEEN any AI. This means you aren’t screwed when one particular model is taken from us. I wrote extensively about this last week https://aiwithkyle.com/ai-news/build-one-shared-ai-vault 

Basically you are building in your own flexibility NOW before you need it.

To the Task,

Kyle