If you’ve been following me for a while, you know I use all the different AI models out there to help my clients choose the best ones for them. They all have different strengths and weaknesses, and those differences don’t always show up in benchmark numbers. For my work, I need models that can create great videos, build powerful presentations, analyze spreadsheets, and produce highly polished Word docs. Almost every client I have still works deeply in the Microsoft Office ecosystem, so I need my tools to work there too.
For the last year, Claude has been the undisputed leader (for me) in creating professional knowledge work documents. Its writing, creativity, and professional formatting in PowerPoint, Word, and Excel have been exceptional. As the models improved (with Opus 5.5 and Fable 5.2), the quality has continued to rise. ChatGPT has been playing catch-up in this space for a while, and while I find their models can do exceptional analysis, I find the professional formatting lacking. Google has been an also-ran in this space for a long time, and they couldn’t even really handle Office documents. That has now changed. With the launch of the new Gemini 4 series - Argon, Carbon, Hydrogen, and now Neon - Google now plays really well with Office documents.
This week we saw the long-awaited launch of GPT-6.2 Astra and GPT-6.2 Sol from OpenAI. Many in the industry (including me) were hoping it would catch up and/or pass the latest Claude models. In my initial testing, it wasn’t even close. I kept telling people how much better Opus and Fable were, and then Gemini 4 Neon joined the party. I feel like Neon is now firmly in the number-two spot for frontier models. But I wanted more than a gut feeling. I wanted something people could tangibly see and test for themselves.
So here’s what I decided to do. I gave Fable 5.2 Extra High in Claude Code in the Claude Desktop app, GPT-6.2 Astra Extra High in Codex in the Codex App, and Gemini 4 Neon Extra High in Antigravity the same set of challenges. I’m having each model do the same four things with the same prompts. I think it’s really useful to compare and contrast how these models do things.
Challenge #1: Create a highly polished video
I’m doing this more and more - creating kickoff videos for clients and showing them what the models are capable of creating. Usually, the videos I do are for company training sessions and have a more corporate look and feel. Given that we’re entering the holiday season, I decided to do something a little different. I also wanted to make sure the models did original research, crafted a story, and came up with something new that I hadn’t done before. This was the starting prompt I gave each model (in a fresh chat session).
Prompt:
I'm doing benchmark testing comparing the latest state of the art models. I live in the city of Westlake Village, CA.
Go research all of the publicly available information you can find about the history of Westlake Village, how the city evolved and what it's like today. Then using blender, davinci, and three.js to make a video, and my eleven labs subscription for any voice you want to use. You can use my ffmpeg to put it all together. Do not use my higgsfield account.
I want you to create a visually stunning and compelling overview for the upcoming holiday season and why people should come visit the city. Use the city colors. I want it fast and jumpy in a modern style also while being festive and friendly. Go create a first version for me. You can also use hyperframes. This will be a one-shot attempt so make your best efforts in this first go-round.
Fable 5.2 Results
In my opinion, this was the strongest and most festive video of the three. It starts with a lot of motion; it has great animations, a good story, a great voice, and really good detail. I even like how they did the ending “Westlake Village” text as holiday lights. Excellent job.
GPT-6.2 Astra Results
To say this video was uninspired, boring, and not festive would be an understatement. It looks like a bad corporate presentation that someone put together, and there was even an audio glitch where they say, “Meet me at the lake. The ” and then nothing. Most of the shots are from summertime, and it includes a lot of photos. This video was a distant third and not one I would ever consider using.
Gemini 4 Neon Results
I like how this starts with really festive music and a holiday look and feel. It’s definitely a strong second (in my opinion) to the Fable video. I like the messaging but did not love the 3-d topographical map it used for a while - I wasn’t really sure why they did that, and it took away from the story. The overall story, voice, and music were great - just a strange choice for a graphic in the middle. I will say that I thought their animated boats and other objects in the last part of the video were superb - even better than what Fable did. Overall, an exceptionally good effort. If it didn’t have the strange topographical map, I probably would’ve tied it with, or maybe even put it a little ahead of, what Fable did.
Challenge #2: Create an exceptionally detailed and highly polished Excel spreadsheet for a fictitious business
This one requires the models to research an industry, think about business strategy, and model the economics for the next five years. It requires building an executive dashboard, income statements, balance sheets, and more. It’s a good test of model-building and analysis, and it requires checking for errors at the end.
What I was looking for in each was the following:
1) Did it successfully research the industry and build a viable spreadsheet with logical assumptions around economics, growth, and more?
2) Did the executive dashboard have solid insights and the kinds of things that someone would actually use to run the business?
3) Were sheets built properly with formulas where there should be formulas?
4) Did it show that it checked its work?
Prompt:
I want you to go build an excel file for me. It’s a five year set of financial projections for a fictitious saas business. The business provides pet insurance and wellness. Go research the idea and the business market. Build a detailed 5 year P&L. There should be an assumptions page, an executive dashboard page that could be used to run the business, income statements, balance sheet and statement of cash flows. Make sure everything that should be formula driven is. Color code things as appropriate. Make this a highly professional spreadsheet with professional polish. One key area to focus on when driving revenue growth is assumptions around routes to market, revenue from those routes, and on the cost side of that you need to build out things like sales teams, partnerships, etc. Be thorough and document your assumptions Results go into a single excel file with as many tabs as needed.
Fable 5.2 Results
Claude built Bowline Pet Health. It built an incredibly detailed Excel spreadsheet with 16 tabs. Claude built the best, most detailed spreadsheet in the group, and it wasn’t even close. Take a look at the Dashboard page. The formatting is perfect (no overlapping text), the charts are readable and logical, and it brings everything I would want to see when running the business onto a single page. The additional pages were all professionally formatted, easy to read, and very logical. I really liked the market research tab and the routes to market. Claude clearly did a ton of thinking building out this model.
GPT-6.2 Astra Results
ChatGPT built out 15 tabs. One look at the executive dashboard, though, made me realize its design skills are the weakest. It had limited tables and only two charts. Definitely not the kind of dashboard you would put in front of an executive team. The detailed pages in the middle (like Membership) were much better designed and very logical. The market notes were good, and while less detailed than Claude’s, they were well organized, which I liked. I thought the “checks” page was really well designed and easy to read.
Gemini 4 Neon Results
I liked the content of the executive dashboard page, but as you can see, Gemini needs to work on formatting. Text overruns on charts; some labels are hard to read, and the dark background lines make them look basic and unpolished (compared to both Fable and Astra). I do like how Neon built out the model’s thinking and assumptions. Some of the interior pages are really well designed, while others have formatting decisions that look a bit odd (particularly on the market research slide). The one piece Neon didn’t include here was a sheet showing the math checks out (both Astra and Fable had it).
Overall, I felt that Claude built what I asked for. A highly detailed, highly polished spreadsheet that I could work with. Both the Astra and Neon spreadsheets would need additional rounds of back-and-forth to reach a good starting point. They were both “good” - but definitely not very good or great.
Challenge #3: Create two highly polished Word documents based on the spreadsheet content—one for a Board of Directors update and the other for Investors.
This one required the models to create professional messaging and formatting for two distinct audiences. It required thinking through the right core messages for each group and how to format them properly.
Prompt:
Now - create two highly polished word documents based on that spreadsheet. The first is an update to the board of directors with these as the latest projections. Talk about the insights, growth in the business, how we will achieve those goals, the risks and include key charts and tables from the spreadsheet. The second word document is an overview for potential investors in our series A round. Create a short and compelling doc about the business, what we’ve achieved so far, where we’re going and why investing in our next round is a smart thing.
Fable 5.2 Results
Fable did an excellent job. The formatting, the content, the structure. It was really well done, easy to read, and its use of highlights was exceptional. The tables were perfectly formatted, and the charts were very well done and easy to read. This is the level of polish I would expect from a very seasoned consultant working at a top-tier firm.
GPT-6.2 Astra Results
The first thing I noticed in the Astra Board update was that it jarringly switched fonts between the first and second paragraphs (for no reason). The headline was in Arial. Then it switched to 12-point Georgia, and the next paragraph switched to 10.5-point Arial. It does this at the top of each page, and it makes no sense. The content also wasn’t in good enough shape for me to consider putting it in front of a Board as is. It definitely needed rework.
Gemini 4 Neon Results
Neon delivered an excellent set of documents! They were very well written, with great structure and a very high level of polish. The tables and charts were easy to read, and the Board document started with a really strong executive summary. Exactly how I would start a document like this. I personally found the Claude document more to my taste in terms of layout and colors. It was a little less dense than the Neon document, but both models created excellent content.
Challenge #4: Create two highly polished Word documents based on the spreadsheet content—one for a Board of Directors update and the other for Investors.
The final challenge was to build a really exciting PowerPoint document from all of this content for the beginning-of-the-year annual Sales Kickoff (SKO). I wanted the models to be creative and deliver something that would bring the right level of energy to an SKO. I wanted each slide to be visually appealing and not constrained by any of my existing templates.
Prompt:
I want you to create a 20 page highly engaging and professional powerpoint presentation for the annual sales kickoff (SKO) meeting around the state of the business. You do not need to use my current template - you are free to design one that fits the theme of the update and the business. Also use hyperframes, three.js, eleven labs, chatgpt image generation, etc- and any other tools you have access to (except higgsfield) to make sure there are great designs and layouts on every page. It should be energetic and make people excited for the upcoming year. Make sure each page is has interesting design elements (while maintaining an overall look and feel to the presentation).
Fable 5.2 Results
Fable’s presentation was just exceptional. The cover slide had great animations when viewed in presentation mode. On page 2, it built a complete video that was fantastic. Great voiceover, great animations, great messaging. It’s exactly the type of content you would want to see in a sales kickoff. The level of attention to detail in this presentation was world-class. It would have taken a team a long time to get here, and I doubt they could have done as well. The colors, the use of images with text overlays, and the animated charts on page 12 were standout elements. The overall flow, messaging, and design were absolutely world-class.
GPT-6.2 Astra Results
This deck felt like it was built by a seasoned strategy consultant with limited design chops. I personally found it flat and uninspiring. It lacked creativity, animations, and a consistent design theme throughout. It wasn’t well positioned for a sales audience attending an SKO, and the messaging was completely off the mark. I would never use this deck.
Gemini 4 Neon Results
Gemini 4 has come a long way in a short time. The deck’s visual design was excellent. The colors, the graphics, the messaging - exactly what you would want in an SKO deck. Like the Word docs it created, these slides are very information-dense. Probably a little too dense for a sales kickoff. Some slides would be better suited for a Board presentation or a company all-hands. Neon also created a video on slide 2, which I was excited to see. I found the voice-over too fast, and the video itself was jerky when it played. It definitely needs additional polish. But the overall look, feel, polish, and messaging make this a strong number two behind Fable 5.2. It’s great to see a Gemini model able to do this kind of Office work with this level of polish.
Final Thoughts
My world revolves around creating highly polished professional content. Claude is just exceptional in this regard (both Opus 5.5 and Fable 5.2). In my daily work, I find it consistently generates much better content overall than any ChatGPT model. I am extremely excited to see how well the new Gemini 4 models are doing, and I plan to use them more and more given what they are capable of. I’m fortunate to have access to world-class models. Not everyone does.
I want to highlight two other use cases. First, these models have become excellent video editors. I’ve had Claude edit my training videos, and it does an incredible job. Astra (v6) is also very good, but I find it a slight half-step behind Claude. I haven’t had a chance to try 6.2 or Neon for video editing yet, but I will. It’s a great use case. The one area where I find the ChatGPT models absolutely superior is computer use. Claude is very good, but the GPT models are just excellent. I haven’t had a chance to test computer use with the latest Gemini 4 models, but I will over the next week or so.
I have put all the files I created in a shared Google directory here so you can download the content and look at it yourself. Link here.
I hope you enjoyed reading this. If you did, please share it with others. Here’s a picture of Magnus and I from the other day.












