❝

TL;DR: GPT-6.2 Astra and Sol dropped on November 17, 2026, and honestly, the internet kind of lost it. Some people called it a disaster. Others liked it. The benchmarks look strong on paper. So what's actually going on?

This article cuts through the noise, checks the numbers, and walks through two real-world tests: a full website build and a 3D-printable pendulum clock.

  • Models released: GPT-6.2 Astra (premium) and GPT-6.2 Sol (budget)

  • Top benchmark: Agents' Last Exam v1.1, Astra leads at 78.3%

  • Biggest community complaint: Writing quality, instruction following, inconsistency

  • Sol's key strength: Near-Astra coding performance at about 4x lower cost

Introduction

OpenAI releases GPT-6.2 Astra and Sol. They're excited. But the internet? Not so much. People were like: "Just send them back. They're not ready." 😂

Chubby called the release a catastrophe. Matt Schumer refused to even write a review. And Theo published "Why I Hate GPT-6.2 Astra", which has already pulled in over 2.1 million views.

why-i-hate-gpt-6-2-astra

The hate is sooo….

Okay, but WHY? Are these models really that bad? Or are people judging them based on hype, expectations, and a few bad experiences?

I got curious, so I decided to test them myself.

In this guide, I'll show you how I tested GPT-6.2 Astra and Sol, what tasks I gave them, where they failed, and where they actually surprised me.

Let's see if the internet is right.

❝

I. Why Is Everyone So Angry About GPT-6.2 Astra & Sol?

So, I need to tell you, the community reaction to GPT-6.2 was something else.

OpenAI dropped Astra 6.2 and Sol 6.2, and almost immediately some of the most well-known voices in AI started posting about how bad their experience was.

Here's a quick rundown of what people actually said:

Who

What they said

Chubby

Called the release a "catastrophe," and went further, saying OpenAI had just handed Anthropic the best possible promotion for its upcoming IPO.

Matt Shumer

Said the models were "essentially unusable" and that he wouldn't even publish a full review because his experience was so bad.

Theo (t3.gg)

Was frustrated enough with Astra 6.2 that he wrote an entire letter to OpenAI detailing the problems. OpenAI fixed a few of the issues he reported, but the changes still weren't enough to satisfy him.

roon

Had a different take. He didn't like Sol 5.6 when it first launched (and it later became a well-loved model), but he actually likes Astra 6.2, even with all this criticism flying around.

Sam Altman

Acknowledged that OpenAI still has "work to do making models nice to work with." That's a pretty direct admission.

So what's the core complaint?

→ Most of the frustration comes from everyday use. People expected GPT-6.2 to make their work easier, but it still struggles with following instructions, writing quality, and handling some surprisingly simple tasks.

But the benchmark results tell a different story. And that gap between benchmarks and real-world feel is exactly what makes this release so confusing.

❝

Learn How to Make AI Work For You!

Transform your AI skills with the AI Fire Academy Premium Plan - FREE for 14 days! Gain instant access to 700+ AI workflows, advanced tutorials, exclusive case studies and unbeatable discounts. No risks, cancel anytime.

II. GPT-6.2 Benchmarks Show Some Impressive Progress

Despite the negative community reaction, GPT-6.2 posts some genuinely strong numbers on professional and scientific benchmarks. Let me walk you through each one, and flag the one that really needs context.

Quick Benchmark Overview:

Benchmark

GPT-6.2 Astra

GPT-6.2 Sol

Context

FrontierCode 2.0

44.2%

43.8%

Fable 5.2 and Opus 5.5 both lead; Sol nearly matches Astra at ~4x lower cost

Agents' Last Exam v1.1

78.3%

73.0%

Astra leads all models

Vals Science Index

68.44%

58.5%

Astra ranks first

ARC-AGI 3

100% / 70.3%

—

Two different configs, see below

1. FrontierCode 2.0: Strong, But Not the Leader

frontiercode-2-0-strong-but-not-the-leader

On FrontierCode 2.0, GPT-6.2 Astra scores 44.2% and Sol comes in right behind at 43.8%, a gap of less than half a point, while costing about 4x less.

So if you're making a purely cost-efficiency decision on coding tasks, Sol is a genuinely interesting option.

One thing to keep in context: neither GPT-6.2 model leads this benchmark. Claude Fable 5.2 is still out front at 49.1%, and Claude Opus 5.5 (45.5% at max effort) also beats both GPT-6.2 models. So GPT-6.2 is competitive on coding.

2. Agents' Last Exam v1.1: A New Lead in Professional Work

agents-last-exam-a-new-high-for-professional-work

This one is legitimately impressive.

Agents' Last Exam v1.1 tests professional-level reasoning across medicine, law, finance, and science, the kind of tasks that actually matter for knowledge workers.

GPT-6.2 Astra takes the top spot at 78.3%, ahead of Gemini 4 Carbon (77.4%), Sol (73.0%), and Claude Fable 5.2 (72.8%). That's a real data point, not just a marketing claim.

3. Vals Science Index: Leading in Scientific Research

vals-science-index-leading-in-scientific-research

GPT-6.2 Astra scores 68.44% on the Vals Science Index and ranks first of the 27 models listed. Sol comes in at 58.5%.

The improvement over the previous GPT-6 Astra is real, still, treat those specific numbers as approximate until Vals publishes the full table.

4. ARC-AGI 3: The "100%" That Needs a Caveat

This is the headline number, and it's the one most likely to be shared. GPT-6.2 Astra is the first model ever to hit 100% on ARC-AGI 3, which is a genuine milestone. But the full result from ARC Prize reads like this:

  • Provider adapter (high-compute) configuration: 100%, at about $4,100 per run

  • Standard harness: 70.3%, at about $24,200 per run

The 100% comes from a specialized high-compute evaluation mode. Both numbers are real and both are verified by ARC Prize. Just make sure you know which one a headline is quoting: the apples-to-apples figure against other models is 70.3%, not 100%.

For reference, ARC Prize also verified Astra at 95.8% on ARC-AGI 2 and 98.5% on ARC-AGI 1.

III. GPT-6.2 Builds a Website from Scratch

Okay, enough benchmarks. Let's see what GPT-6.2 actually does on a real project.

One of the best ways to test a model's coding ability is to give it a detailed, specific brief and see how close it gets to a finished product. So I asked GPT-6.2 to build a complete website for a premium coffee brand called Morning Peak.

Here's the full prompt you can copy and try yourself:

Act as a senior web designer and full-stack developer.

Your task is to build a complete, production-quality website for premium coffee brand called "Morning Peak".

PROJECT GOAL:
Create a modern website that introduces the brand, showcases its coffee products, and encourages visitors to make a purchase.

BRAND IDENTITY:
- Brand name: Morning Peak
- Industry: Premium specialty coffee
- Target audience: Young professionals and coffee lovers
- Brand personality: Modern, warm, minimal, and premium
- Color palette: Warm cream, dark brown, coffee beige, and muted orange
- Typography: Use elegant headings and clean, readable body text.

WEBSITE STRUCTURE:

1. Homepage
- Create a strong hero section with a large coffee image.
- Add the headline "A Better Start to Every Morning".
- Include a short brand introduction and a clear Shop Now button.

2. About Section
- Introduce the brand and its coffee philosophy.
- Use images and text in a balanced layout.

3. Product Collection
- Display at least 6 coffee products.
- Each product must include an image, name, price, and short description.
- Add functional Add to Cart buttons.

4. Why Choose Us
- Highlight the brand's coffee quality and sourcing process.
- Use simple icons and short descriptions.

5. Customer Reviews
- Display customer testimonials in a clean layout.

6. Newsletter Section
- Add an email subscription form with basic input validation.

7. Footer
- Include navigation links, social media links, and contact information.

FUNCTIONAL REQUIREMENTS:
- Make the website fully responsive.
- Add a working shopping cart with item quantity controls.
- Update the cart total when products are added or removed.
- Make all navigation links and buttons work.
- Add smooth scrolling between sections.
- Include basic form validation.
- Use reusable components and clean code.

DESIGN REQUIREMENTS:
- Avoid generic AI-generated website layouts.
- Use consistent spacing, typography, and visual hierarchy.
- Make the hero section visually impressive.
- Use high-quality coffee images.
- Keep the design clean and easy to navigate.

TECHNICAL REQUIREMENTS:
- Use React and Tailwind CSS.
- Organize the code into reusable components.
- Make sure the project runs without errors.
- Don't leave placeholder sections or unfinished buttons.

Before building:
1. Review the complete requirements.
2. Create a short implementation plan.
3. Identify the main components and their relationships.

Then build the complete website.

After implementation:
- Run the project.
- Test the navigation and shopping cart.
- Check the responsive layout.
- Fix any errors you find.

Don't stop at a design mockup. Deliver a working website.

The result looks pretty good! GPT-6.2 created a complete coffee brand website with a warm color palette, elegant typography, and a clean layout.

The hero section immediately catches your attention, while the product collection makes it easy to browse all six coffee products.

gpt-6-2-builds-a-website-from-scratch

I especially like how the design feels consistent across the page. The coffee images, product cards, and small details all fit the Morning Peak brand. It looks much closer to a real coffee business website than a basic AI-generated landing page.

Of course, a good-looking website is only part of the test. I'd also check if the shopping cart works, the navigation links go to the right sections, and the layout looks good on mobile. These tests will help us see how well GPT-6.2 handles the actual development work, beyond the visual design.

IV. GPT-6.2 Builds a 3D-Printable Pendulum Clock

After testing GPT-6.2's website-building skills, I want to give it a much harder challenge: designing a 3D pendulum clock that can be assembled and actually work.

For this test, I'll ask GPT-6.2 to design each part using CAD, check the dimensions, and make sure all components fit together. Then, I'll use Python to test the mechanical movement and Blender to simulate the complete clock.

You can copy the prompt below and try it with GPT-6.2.

Act as a senior mechanical engineer, CAD designer, and Python developer.

Your task is to design a fully functional, 3D-printable pendulum clock from scratch.

PROJECT GOAL:
Create a working mechanical pendulum clock with individual 3D-printable components. The final design must be suitable for assembly and include a simulation of its mechanical movement.

DESIGN REQUIREMENTS:

1. Overall Design
- Create a compact desktop pendulum clock.
- Use a clean, modern mechanical design.
- Include a visible pendulum and clock mechanism.
- Make the structure stable enough to stand on a desk.
- Keep all components suitable for standard FDM 3D printing.

2. Mechanical Components
Design all necessary components, including:
- Main frame and support structure.
- Pendulum rod and bob.
- Gear train.
- Escapement mechanism.
- Clock hands and dial.
- Axles, connectors, and mounting parts.

3. CAD DESIGN
- Create each component as a separate CAD model.
- Use precise dimensions and define the relationship between connected parts.
- Check gear alignment and mechanical clearances.
- Make sure all components fit together during assembly.
- Use parametric modeling where possible.

4. MECHANICAL VALIDATION
- Calculate the pendulum length and expected oscillation period.
- Check gear ratios and rotational movement.
- Write Python scripts to test the mechanical relationships.
- Identify potential collisions or parts that may prevent movement.
- Fix design problems before finalizing the models.

5. SIMULATION
- Import the designed components into Blender.
- Create an assembly of the complete clock.
- Simulate the pendulum swinging.
- Show how the gear train and clock hands move.
- Demonstrate the complete mechanism in motion.

6. FINAL FILES
Deliver:
- Individual 3D models for every component.
- Complete assembled CAD model.
- STL files ready for 3D printing.
- Python validation scripts.
- Blender simulation file.
- A complete assembly guide with numbered steps.
- A parts list with dimensions and quantities.

WORKFLOW:
Step 1: Analyze the mechanical requirements and create a design plan.
Step 2: Design the mechanical system and calculate the dimensions.
Step 3: Create individual CAD components.
Step 4: Assemble the components and check compatibility.
Step 5: Run mechanical calculations and collision checks.
Step 6: Simulate the clock movement in Blender.
Step 7: Export the final 3D-printable files and assembly instructions.

IMPORTANT RULES:
- Do not create a decorative clock that only looks functional.
- Every mechanical component must have a clear purpose.
- Do not claim the clock works unless its mechanical behavior has been validated.
- If a design problem appears, revise the affected components.
- Keep all design files organized and reproducible.

Before starting, inspect the available tools and choose the best CAD and simulation workflow.

First, present your design plan and wait for approval. Then build the complete project.

The main thing I want to check is whether GPT-6.2 can handle the full design process. A clock may look impressive in a render, but the real challenge is making sure all the gears fit, the parts can be assembled, and the pendulum can move as expected.

gpt-6-2-builds-a-3d-printable-pendulum-clock

I'll also look at how GPT-6.2 designs individual components and connects them into one complete mechanical system. The exploded view helps us see how the different parts fit together, while the close-up shows the level of detail in the gear mechanism.

exploded-cad-view-of-clock-components-with-close-up-gear-mechanism

Finally, I'll use the Blender simulation to check how the pendulum and gears move together. If GPT-6.2 can create the CAD files, STL models, Python scripts, and Blender simulation, it'll be a strong example of its engineering capabilities.

However, we'll still need to 3D-print and assemble the clock to confirm that the mechanism actually works.

V. Should You Use GPT-6.2? Final Verdict

After testing GPT-6.2, I think it has some interesting strengths, especially in website building and 3D engineering projects. However, I'd still compare it with other models before making it my main choice.

Here's my quick breakdown:

Task

My recommendation

🌐 Website development

Worth testing GPT-6.2

🧩 3D modeling and CAD design

Worth exploring GPT-6.2

⚙️ Engineering tasks

Test its CAD, Python, and Blender capabilities

💻 Coding

Compare with Claude Opus 5.5 and Fable 5.2

✍️ Writing

Compare output quality and instruction following

🏁 My final take: GPT-6.2 is worth trying, especially for complex projects. However, impressive demos and benchmark scores don't always mean reliable results. I'd choose the model that performs best on the tasks I actually need.

Pricing quick reference:

Model

Input (per 1M tokens)

Output (per 1M tokens)

Note

GPT-6.2 Astra

$8.00

$40.00

~20% cheaper than GPT-6 Astra

GPT-6.2 Sol

$2.00

$10.00

~4x cheaper than Astra 6.2

Don't judge GPT-6.2 by the hype or the backlash. The community reaction is real feedback about everyday usability, but the benchmark numbers and project demos show genuine capability on complex technical work.

Try it on something you actually need to do, compare it against the model you're currently using, and let the results decide.

Conclusion

🚀 GPT-6.2 has given us an interesting release, but the community's mixed reactions tell a bigger story. Its benchmark results look impressive, and our website and 3D clock projects show some exciting possibilities beyond traditional coding.

However, a great demo doesn't always mean a reliable model. You still need to test how well GPT-6.2 follows instructions, handles complex tasks, and delivers working results.

If you are interested in other topics and how AI is transforming different aspects of our lives or even in making money using AI with more detailed, step-by-step guidance, you can find our other articles here:

Reply

Avatar

or to participate