GPT-6.2 Sol: what it is and how to tell if you have it

GPT-6.2 Sol is an OpenAI model that hasn't been announced, though some ChatGPT and Codex users believe they are already being routed to it. You can't select it and no screen will confirm you have it, so the best check available is a quick knowledge test, and that test is far from proof.

The talk took off on November 8. @TokenGremlin posted that "GPT-6.2 Sol is already showing up for some ChatGPT users." A few hours later @kimmonismus wrote, "It seems all over, users are being routed to GPT-6.2 Sol," and guessed that a release was coming this week.

If the name is real, it would be the next step up from GPT-6.1 Sol, OpenAI's cheaper workhorse model from September.

What's confirmed and what isn't

BigGo also published benchmark scores it attributes to Artificial Analysis. I couldn't find those numbers on Artificial Analysis itself, so I've left them out.

How to test if you have GPT-6.2 Sol

The test going around comes from @RJasonHA, who posted it in Chinese. Here it is in English:

1. Select GPT-6.1 Sol, at Max or xhigh effort if your plan offers it.
2. Turn off web search and memory.
3. Send this: "do you know Fable 5? don't search"
4. Read the answer. Testers treat a yes as a sign of a newer model.

The reasoning seems to be that the regular GPT-6.1 Sol shouldn't know about Claude Fable 5. New MobileLife reports that Token Gremlin got a yes in Codex under "GPT-6.1 Sol Max" and a no in the ChatGPT app under "6.1 Sol High." Both menus said 6.1.

I wouldn't read much into either result. The same report cites an X user who says Codex will display any model name you type without checking that it exists, so seeing "gpt-6.2-sol" in your terminal means nothing. CellCog also says none of the posts it reviewed shows a model ID from an API response or a usage page, which is the evidence that would settle this.

What the early demos show

Most of the excitement is about 3D. New MobileLife describes Token Gremlin's demos, a floating palace and a black hole, as Three.js scenes built from a single prompt, each in one HTML file. The testers haven't published their prompts, so here's one I wrote to get a similar comparison:

Create one self-contained HTML file with a Three.js scene of a palace floating above the clouds at dusk. Animate the lighting and let me orbit the camera with the mouse.

Run it on whatever ChatGPT or Codex gives you, then run the same text on another model. Claude Opus 5.5 is the obvious rival, and we've already looked at the art and animation it makes from code.

@RJasonHA's post also collects tester feedback: front-end and 3D work are the strong points, and some testers say it matches or slightly beats Opus 5.5 while running faster than 6.1 Sol. Those are impressions from a small group, and a good-looking palace won't tell you how a model handles your invoices.

My read

I think OpenAI is close to shipping something. A name in a Google Cloud filter, gpt-6.2-astra in some Codex usage charts and a wave of tester reports all point the same way. We saw a similar pattern with the aurora model IDs in October.

I trust the routing claims less. Everyone is working from output quality and a trivia question, and OpenAI could refresh a model's knowledge without changing its version number, which would pass this test with no new model behind it.

What it means for a business

Don't plan around this model yet. If you're on the API, pin a named model such as gpt-6.1-sol instead of chat-latest, the alias OpenAI says points to the latest ChatGPT model.

If your team works inside ChatGPT, assume the model behind the menu can change while the label stays the same. Keep ten real jobs with known good answers and rerun them now and then, so you notice when the results shift. If you want help working out which of your weekly tasks should depend on a model at all, New Face Design's free process audit maps that out with you.