0x Insider

I test new AI models on real work and post what actually happened. Failures included.

[0x Insider @0xInsiderf5]

You pay for Astra. Some of you are getting Luna and the menu still says Astra.

One Pro 20x user logged nine hours. 247 responses carried a model field. All 247 said gpt-5.6-luna.

Same account, same prompt, eleven minutes apart, two different models.

The only tell is a header length. 292 characters, you get what you picked. 312, you don't.

Nothing in the app tells you.

[0x Insider @0xInsiderf5]

@BEJLES777 still ticking, almost 35 hours now. no idea what happens, so we're leaving it open to find out. will post the number tomorrow night

[0x Insider @0xInsiderf5]

🚨 ChatGPT showed GPT-6.1 Sol "Working" on our video for 33 hours. Its cloud machine had died at hour 2.5.

All 15 finished frames went with it. Claude Sonnet 5.5 finished the same prompt in 7h 35m.

How long would you have waited?

[0x Insider @0xInsiderf5]

chatgpt's "working" timer is still running right now, 34 hours and counting. the rest is from sol's own postmortem. the machine went offline at hour 2.5, cause unknown, so the trigger wasn't the model. but the scene and all 15 frames lived only on that machine, and openai's own support assistant says long runs should save to the library as they go. to its credit, sol graded its own deliverable failed and said i didn't cause it. sonnet ran on max, sol on medium, the highest our plan allows. full sonnet run in https://t.co/nYqF0A3wyT

[0x Insider @0xInsiderf5]

@bubosees fair, missed those. where are the jev runs, would love to see them next to these three

[0x Insider @0xInsiderf5]

@bubosees add jev to the next one. most of these claim they beat jev on their own tests, a snake board would settle it

[0x Insider @0xInsiderf5]

@bubosees watching it now. funny thing is all three dropped on oct 1 and this is the first real head to head i've seen

[0x Insider @0xInsiderf5]

Update, two days later. The account behind it renamed itself from claudesfun to codersrh, display name "coders". Same claim post, and the old handle now leads nowhere.

[0x Insider @0xInsiderf5]

Anthropic launched claude[.]dev yesterday. This is claudes[.]dev, one letter longer, and it is not Anthropic.

Anthropic has no token. This account joined in September, follows 2 people, and its CLAUDES claim post hit 647K views.

Check the letters before you connect a wallet.

[0x Insider @0xInsiderf5]

https://t.co/bGswNTJzsz, it checks OpenAI's Codex source every 15 minutes. This one came in with PR 49318.

[0x Insider @0xInsiderf5]

OpenAI just reshuffled the model menu in Codex. GPT-6.1 Sol now sits above Astra, and the old workhorse GPT-6 Sol is officially "previous generation".

Our tracker caught it in the public source this morning.

[0x Insider @0xInsiderf5]

@TheGeorgePu curious how you measured it. gpt-6 sol finished our film test in 9 minutes, then we found 47% of its frames never move. sometimes efficient just means skipping the work

[0x Insider @0xInsiderf5]

@TomSnrub @TheGeorgePu fair, 256mb was half the iphone 4 and the back camera was under 1mp. but for a kid who couldn't get an iphone it was the closest thing, which was george's whole point

[0x Insider @0xInsiderf5]

We asked Claude Sonnet 5.5 for the most realistic video it could make. Max effort, no time limit. This is its ceiling.

No camera, no video model, no stock 3D. Every frame and every sound came out of code it wrote.

First it checked the machine. One CPU core, 3.9 GB of RAM, no GPU. It timed a test frame and did the math itself. 8 samples per pixel would have taken about 12 hours, so it went down to 2 and let a denoiser clean up the noise.

It kept the whole city static to save 3.4 seconds a frame and moved the camera through it instead. Rain, steam and lightning were painted on top afterwards.

The sandbox rebooted at least three times mid-render and killed the job each time. It wrote a watchdog, but the render only moved while the model was mid-turn. 7h 35m on the clock, 3h 09m of that sitting idle.

The soundtrack is code too. It picked 128 BPM so every hit lands on

[0x Insider @0xInsiderf5]

@rwspatin @theo @Railway to be fair those were different jobs, a cipher search vs 360 rendered frames, so 53 min isn't the same work. but the vm point stands. i want to run the same render on our own box and post the clock both ways

[0x Insider @0xInsiderf5]

@rwspatin @theo @Railway that's the move. ultracode did it on its own in claude code, split my enigma run into 2 subagents and finished in 53 min. the app is the problem, its vm rebooted at least 3 times mid render and the job only moved while the model was mid turn

[0x Insider @0xInsiderf5]

@rwspatin @theo @Railway tools run on their own, it never asks. the stops are per turn caps, max length and a tool use limit, then it waits for continue. and it's worse than 20%, the final log says 3h 09m idle out of 7h 35m

[0x Insider @0xInsiderf5]

@TheGeorgePu 246gb up front is refreshing. this week i watched a model render 15 seconds of film on 1 vcpu and 3.9gb ram, so i get why you said it first. what does the 123b do on mle work that the 30b can't

[0x Insider @0xInsiderf5]

The real one is claude[.]dev/terminal, where people are getting plushies and stickers. The fake one wants you to finish steps so 200,000 tokens get reserved for your wallet.