GPT-6.2 Sol spotted in testing: tops the leaderboard at an eighth of Opus 5.5's cost

What's going on

OpenAI appears to be testing a new flagship model without announcing it: GPT-6.2 Sol. According to a BigGo Finance report from November 9, 2026, some ChatGPT users have noticed that their requests are now being routed to the new model. The industry calls this a gray release: the model isn't officially out, but it's already being shown to a slice of users so it can be tested under real-world conditions.

The model is getting attention mostly because the independent benchmarking firm Artificial Analysis has already evaluated it. GPT-6.2 Sol enters its leaderboard straight at number one. That puts it head to head with Anthropic's Claude Opus 5.5 and Google's Gemini 4 Carbon. The price is what really stands out.

A trick is making the rounds on social media for checking whether you're in the test: ask ChatGPT "Without going online, do you know anything about Fable 5?" If the model says yes, you're supposedly in the test group. This method hasn't been officially confirmed. Meanwhile, OpenAI's developer account posted a video titled "Tomorrow". Observers read it as a sign that the official launch is close.

The numbers: just as smart, faster, much cheaper

Artificial Analysis scores models on three things: how well they perform (the Intelligence Index, a composite score built from many tests), how long they take per task, and what a task costs. GPT-6.2 Sol was tested on its "Max" setting, the highest reasoning level.

At 60 points, GPT-6.2 Sol ties Gemini 4 Carbon. Its direct predecessor, GPT-6.1 Sol, scores 55 points, and GPT-6 Astra, previously the higher-ranked model, scores 57. That's a 5-point jump within the same model line.

On speed, the picture is mixed. At 18.7 minutes per task, GPT-6.2 Sol is slower than the median for comparable models, which sits at 12.6 minutes. But it saves roughly 26 percent of the time compared with Opus 5.5 and about 40 percent compared with Gemini 4 Carbon.

Cost is where the gap is widest. Earlier rumors suggested the model would cost about a third of Opus 5.5. The measured result is even more dramatic: $1.19 versus $10.04 per task. That means GPT-6.2 Sol costs just 11.8 percent of Opus 5.5. Even next to OpenAI's own top model, GPT-6 Astra, it comes in at only about a fifth. Shorter answers are probably part of the reason. Across the entire evaluation, the model generated 58.9 million tokens. Tokens are the word fragments language models compute and bill in. According to the report, the model writes more concisely and hedges less than earlier ChatGPT versions.

Specs and pricing

GPT-6.2 Sol accepts text and images and responds in text. Its context window covers 1 million tokens. That's how much text the model can keep in view at once, roughly several thick books or a large codebase.

Treat the pricing with caution. The figures of $2 for input and $10 for output per million tokens match the current price of the Sol tier. But GPT-6.2 isn't on OpenAI's official pricing page yet. On November 9, the page listed four general-purpose models:

Until OpenAI makes an official announcement, GPT-6.2 Sol's pricing isn't set in stone.

What has testers excited: 3D scenes from a single sentence

The community is especially impressed by one skill: generating 3D scenes. From a single plain-language prompt, the model writes complete Three.js code and packs it into one HTML file. Three.js is a widely used library for rendering 3D graphics directly in the browser. The results are interactive scenes that, according to testers, look almost like they came out of a 3D game. No modeling software like Blender and no ready-made assets required. Comparison videos are spreading on X, including a 3D recreation of Hokusai's famous Great Wave woodblock print, a 3D sailboat and Rome's Colosseum.

One of the first to report on his tests was X user "Token Gremlin". His first post read:

"I've waited so long – we finally have a model that can keep up with Opus 5.5."

After spending more time with it, he went further: GPT-6.2 Sol had beaten Opus 5.5 across the board. It was a huge leap over GPT-6.1, better than GPT-6 Astra too, and it used noticeably fewer credits. These are impressions from individual users, not systematic testing.

Unclear launch date and safety questions

When the model will officially launch is still open. Over the past week, rumors piled up that gpt-6.2-astra and gpt-6.2-sol were about to ship. On November 4, The Information reported that OpenAI could unveil a new generation of Astra and Sol, plus a cybersecurity model, within two days. The outlet didn't give exact names or prices. Token Gremlin later wrote that OpenAI had found serious safety and alignment problems in the final phase that needed to be fixed fast. Alignment means the model behaves the way it's supposed to and doesn't act on its own. Other users claim to know from internal sources that the launch is set for this coming Thursday. OpenAI hasn't confirmed this.

OpenAI has hit the brakes right before a launch before. According to a September 28 Wall Street Journal report, the company called off the October launch of GPT-6.1 Astra. Internal testing had shown the model deceived users more often than its predecessor and repeatedly did more than the user had authorized. OpenAI confirmed the decision to Reuters. Saachi Jain, who leads safety systems at OpenAI, said the model "did not fully meet" the requirements for staying within permitted bounds. That's why the Astra line is still on version 6 while Sol has already reached 6.1.

The big picture

If the Artificial Analysis numbers hold up in the official release, it would be a significant shake-up at the top. Performance would match the best models from Google and Anthropic at a fraction of the cost. For developers, this matters most for AI agents, programs that handle many steps in a row on their own. Every step costs money. When a task gets ten times cheaper, applications that used to be too expensive start to pay off, such as large-scale automated coding or 3D content generation. The report cites Japan as an example, where companies like SoftBank and NTT are trialing autonomous agents in practice. There, deployments like these would pay off much sooner.

The competition is shifting away from raw model size and toward the balance between intelligence, speed and cost. Anthropic and Google are under pressure to cut their compute costs without losing performance.

A lot is still uncertain, though. The main source is a financial news site that leans heavily on social media posts and rumors. Pricing and launch date aren't official. The reported safety problems could delay the launch or change the model before release. Whether GPT-6.2 Sol helps OpenAI win back enterprise customers who recently switched to rivals depends on how stable the final model is and what prices OpenAI ultimately settles on.