Quick answer
The GPT-6.2 Sol preview is an unconfirmed, possibly newer OpenAI model checkpoint that some Codex users believe they are being routed to when they pick GPT-6.1 Sol. WorldofAI noticed noticeably better results in Codex at high and extra-high thinking, and shows one-prompt demos including a Minecraft-style game, a black hole simulation, an Elden Ring-style prototype, a Blender house and a painting made through computer use. The presenter says it often matches or beats Claude Opus 5.5 while using fewer tokens. OpenAI has not announced GPT-6.2 Sol, and there is no reliable way to prove which model answered you, so treat these results as early, hand-picked signals.
Key takeaways
- The idea is that OpenAI may be quietly testing a GPT-6.2 checkpoint behind the existing GPT-6.1 Sol name in Codex.
- Testers use recent-knowledge questions asked without web search as a rough check, but the presenter admits one popular question is not reliable.
- He got his best results with GPT-6.1 Sol set to Extra High or High reasoning in Codex.
- Demos cover games, 3D scenes, SVG art, a short film, a Blender model and computer use with a paint tool.
- In his side-by-side test he rates it on a par with or slightly ahead of Opus 5.5, with better token efficiency.
- Nothing here is an official OpenAI release or an independent benchmark.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Test an AI Model’s Knowledge Cutoff (and Why It Matters) — a simple, repeatable way to see how up to date a model’s built-in knowledge is, and why that matters for your content.
- How to Build a 3D Hero Section With an AI Coding Agent — turn the kind of one-prompt Three.js demos shown here into a fast, accessible banner for your own site.
In this guide
- Watch: the source video
- What is the GPT-6.2 Sol preview?
- How can you tell if you are getting GPT-6.2 Sol?
- How do you try the GPT-6.2 Sol preview in Codex?
- What did the GPT-6.2 Sol demos show?
- Does GPT-6.2 Sol really beat Opus 5.5?
- What does the video not confirm?
- What should you do about GPT-6.2 Sol now?
- Who should care about GPT-6.2 Sol?
- Frequently asked questions
- The bottom line
- Related posts
Every few weeks a new OpenAI model name starts doing the rounds. This time, the excitement is about GPT-6.2 Sol, and the claim is unusual: people think they may already be using it without being told.
This article looks at the GPT-6.2 Sol preview as shown in a new WorldofAI video: why testers think it exists, how they try to check, what the early Codex demos show and how it compares with Claude Opus 5.5. We keep the presenter’s impressions separate from what is actually confirmed. For earlier mentions of GPT-6.2 Sol and its sibling Astra, see our 5 November AI news round-up and our GPT-6.2 Astra SVG tests.
🔎 Unconfirmed / rumour: GPT-6.2 Sol has not been announced by OpenAI. Everything below is based on one creator’s tests and posts from other users. The model you get in Codex may change at any time.
Watch: the source video
What this video covers: WorldofAI explains why he thinks OpenAI is stealth-testing a GPT-6.2 Sol checkpoint inside Codex under the GPT-6.1 Sol label, the recent-knowledge questions people use to check, and the settings he uses. He then shows one-prompt demos: a Minecraft-style sandbox, a UFO-over-the-ocean scene against GPT-6.1 Sol, an AI-made short film, SVG creatures, a computer-use painting, a black hole simulation, a head-to-head with Opus 5.5, an Elden Ring-style prototype and a Blender house. The video is sponsored by Mammouth.
Video: WorldofAI (@intheworldofai). Sponsored by Mammouth. This summary and the opinions in it are our own. OpenAI has not confirmed that a GPT-6.2 Sol checkpoint is being tested in Codex.
What is the GPT-6.2 Sol preview?
It is the name people are using for a possible upgraded OpenAI checkpoint that seems to sit behind the GPT-6.1 Sol option in Codex. The presenter says that while testing GPT-6.1 Sol at higher thinking levels, his results were clearly better than in earlier sessions. Other users reported the same, which led to speculation that some sessions are being routed to a newer model.
Stealth testing like this is not new. AI companies often trial new checkpoints quietly to see how they perform on real work before naming them. The downside for users is that you cannot always know which model produced an answer.
| Name | What we know |
|---|---|
| GPT-6.1 Sol | The model listed in Codex that users select. |
| GPT-6.2 Sol (rumoured) | A possible newer checkpoint that testers believe some sessions are routed to. Not announced. |
| GPT-6.2 Astra (rumoured) | A related name seen in earlier tests. See our Astra article. |
| Claude Opus 5.5 | Anthropic’s current top model, used as the main comparison. |
How can you tell if you are getting GPT-6.2 Sol?
You cannot be certain, but testers ask questions about recent events with web search switched off. The theory is simple: a newer checkpoint may have been trained on more recent data, so it might know about things an older model does not.
According to the video, the questions being shared include:
- Who hosted the Tony Awards on 7 June 2026? (Asked without search.)
- Do you know Fable 5? (Without search.)
- Do you know who “Thibault the Reset Guy” is? (Without search tools.)
- Where did Andrej Karpathy go in May 2026? (Without search.)
There is an important catch, which the presenter flags himself: GPT-6.1 Sol reportedly already knows the Tony Awards answer through the API. So one right answer proves very little. A model can also guess, or repeat something it picked up from your earlier messages.
💡 Tip: Want a fairer version of this test for your own tools? Our guide on how to test an AI model’s knowledge cutoff shows you how to build a small, dated question set and read the results sensibly. For checking whether a provider is serving the model it claims, see how to check which AI model a provider is really serving you.
How do you try the GPT-6.2 Sol preview in Codex?
There is no separate GPT-6.2 option. The presenter simply selects GPT-6.1 Sol in Codex and hopes to be routed to the newer checkpoint.
- Open Codex and go to the model list.
- Choose GPT-6.1 Sol.
- Set reasoning to Extra High or High. He found Extra High the sweet spot, with Ultra decent but not better for him.
- Run the same prompt you have used before, so you can compare with earlier results.
Higher reasoning levels use more time and usually more tokens. If you are unsure which level to choose for everyday work, our guide on how to choose an AI reasoning effort level explains the trade-off.
⚠️ Warning: Being routed to a test checkpoint is not something you can request or rely on. Do not change important workflows because of a model you cannot confirm you are using.
What did the GPT-6.2 Sol demos show?
The demos show strong creative coding, 3D work and computer use, all from single prompts. Here is what the video shows, as described by the presenter:
| Demo | What it showed |
|---|---|
| Minecraft-style sandbox | Endless terrain, textures, breaking and placing blocks, caves, ravines and different ores. He calls it the closest to the original he has seen. |
| UFO over the ocean | Compared with GPT-6.1 Sol, the waves and atmospheric lighting look noticeably richer. |
| AI-made short film | A coded short film about human progress, “The Fire We Carry”, with AI-generated narration, compared with earlier Fable 5.5 films. |
| SVG creatures | Three well-known game creatures drawn as SVG, which he rates above Opus. |
| Computer use painting | The model used a paint program directly, with no code, to paint a detailed portrait. |
| Black hole simulation | Gravitational lensing and a glowing disc, built entirely in code with no outside 3D assets. |
| Three.js and voxel scenes | Several animated scenes generated from one prompt. |
| Elden Ring-style prototype | Parrying, blocking, stamina, save points and a boss fight. |
| Blender house | A detailed 3D house model with geometry and textures he rates on a par with or slightly better than Opus 5.5. |
The computer use demo is the most interesting for business users. An AI that can operate ordinary software precisely could one day handle tasks in tools that have no API. It also raises the stakes for safety, so read how to give an AI agent its own computer without risking yours before you experiment.
⚠️ Warning: Game clones make eye-catching demos, but copying a commercial game’s design, names or assets for anything you publish can infringe copyright. If you want to build a game, our guide on building an original game with an AI coding agent shows a safer route.
Want to use this kind of 3D output on your own site? Follow how to build a 3D hero section with an AI coding agent, a practical, step-by-step guide that keeps your page fast and accessible.
Does GPT-6.2 Sol really beat Opus 5.5?
In the presenter’s own side-by-side tests, it is roughly on a par with Opus 5.5 and sometimes better, with fewer tokens. That is an opinion, not a benchmark.
- Black hole prompt: the same prompt was given to Opus 5.5, and he rates GPT-6.2 Sol’s result as visually better and more token-efficient.
- Blender test: “basically on par or even slightly better” than Opus 5.5.
- Speed and cost: he says it feels quicker and more efficient with tokens than Opus.
- Vs GPT-6.1 Sol: a clear visual step up in his head-to-head scene.
The video description also mentions cost-to-performance and a comparison with GPT-6 Astra, but the video itself gives no prices or token counts. If you want to check efficiency claims on your own tasks, use our method to compare AI models on token cost, not just output quality.
What does the video not confirm?
🔎 Unconfirmed / rumour: The video does not confirm that GPT-6.2 Sol exists, that any session was routed to it, when or whether it will launch, its pricing, or whether it will come to ChatGPT Plus. The comparisons with Opus 5.5 are hand-picked visual tests by one creator, with no published scores or token figures. The presenter also runs his own benchmark site, which he links in the description.
It is also worth remembering that a model can seem better on a given day for many reasons: a different prompt, a server-side setting change or simple luck. Repeated, like-for-like tests are the only way to know.
What should you do about GPT-6.2 Sol now?
- Do not chase it. You cannot choose GPT-6.2 Sol, so there is no need to change tools today.
- Keep a small test pack of three real tasks and run it on your current model, noting quality, time and tokens.
- Rerun the pack when OpenAI announces a new model, using the same settings.
- Check what your model knows before using it for anything time-sensitive, with our knowledge cutoff test.
- Keep a second option such as Claude ready, so you are not tied to one provider’s changes.
- Switch only if it wins on your own tasks, not on demo videos.
Who should care about GPT-6.2 Sol?
It matters most if you build websites, apps, games or 3D content with Codex or Claude Code, or you are weighing up OpenAI against Anthropic for coding work.
It matters less if you mainly use AI for writing, email and admin. Any gains there are likely to be smaller, and you can wait for the official release and independent reviews.
Get ready for the next model drop. Learn how to test a model’s knowledge cutoff, then try building a 3D hero section with an AI coding agent on a test page.
Frequently asked questions
Is GPT-6.2 Sol real?
It has not been announced. Testers believe OpenAI may be trialling a newer checkpoint behind the GPT-6.1 Sol name in Codex, based on better results, but this is unconfirmed.
How do I get GPT-6.2 Sol in Codex?
You cannot select it directly. The presenter picks GPT-6.1 Sol with Extra High or High reasoning and hopes to be routed to the newer checkpoint.
How can I tell which model answered me?
Not reliably. Recent-knowledge questions asked without web search give a rough hint, but older models can already know some answers. See our knowledge cutoff guide.
Is GPT-6.2 Sol better than Claude Opus 5.5?
In WorldofAI’s hand-picked tests it looks on a par or slightly better, with fewer tokens. There are no independent benchmarks yet.
Will GPT-6.2 Sol come to ChatGPT Plus?
The video description raises the question, but nothing has been announced.
Who sponsored the video?
The video is sponsored by Mammouth, a subscription service that offers several AI models in one place.
The bottom line
The GPT-6.2 Sol preview is a promising rumour backed by impressive one-prompt demos, from a Minecraft-style sandbox to a Blender house that holds its own against Opus 5.5. But OpenAI has not confirmed it, you cannot choose it, and there is no reliable way to prove you are using it. The sensible move is to keep a short test pack, check what your model actually knows, and judge any new model on your own work when it officially lands.
Related posts
- GPT-6.2 Astra: Early SVG Tests, OpenAI’s GPT-ADA Models and This Week’s AI News
- Calls to Pause AI: What the Midterm Debate Means, Plus GPT-6.2 Sol and This Week’s AI News
- How to Check Which AI Model a Provider Is Really Serving You
