How Anthropic's models continue to pass my Hour Glass Digital Twin, this time with Fable 5.1

Continuing my benchmarking of how well AI can build a digital twin of an hourglass timer - this time with Claude Fable 5.1 High using Claude Code app. Compared to Gemini 3.8 and GPT 6 Astra, Fable continues to outshine and perform the best in one-shot building. Sure, there's some refinements to make with a more realistic pass through of the sands at the neck choke point, but if you watched my earlier videos, you can see for yourself how much better Fable 5.1 is when compared to other models. Previous Anthropic models like Opus 4.8 & Fable 5 also did really well, which leads me to conclude that Anthropic's models are better at general intelligence, science and physics models and software coding. Unlike GPT 6 Astra and Gemini 3.8, I have shared Fable 5.1's code on Github here: 

It is nevertheless interesting how these latest models from various companies claim high scores and top-place  in AI benchmark tests. Yes, these models have significantly improved over time, and can do amazing things much greater than an hourglass digital twin like 3D games, world simulations, etc -- yet at the same time struggle to implement my rather simple benchmark test of building a real-world digital representation of an hour-glass! Fable 5.1, Fable 5 and Opus 4.8 all pass my benchmark test, which is also interesting because I do have optimality when it comes to choosing models for cost/performance/quality. My default model for Claude Co-Work and Claude Code is still Opus 4.8, and I'm not sure I'll be willing to switch to Fable 5.1 for all my workflows, the cost is the biggest factor. Having said that, I use Codex Sol as my main coder, pairing with Claude Opus to do my thinking, research, reviews and have a workflow setup where Codex collaborates with Claude. This is because I have far more headroom with OpenAI models than I do with Anthropic, a gap that Anthropic needs to close down soon if they want to people to never leave their platform.

GitHub project: https://github.com/khanmjk/Fable51_Hourglass
Live app: https://khanmjk.github.io/Fable51_Hourglass/

I use this standard prompt for all my tests:

Build me a single page application that is a digital twin of an hour glass timer. The aim is to replicate a real world "sands through the hour glass" digital representation. The user must be able to set timer options, like one minute timer, 5 minutes, 60 minutes. The hour glass must be filled with sand grains, when the timer starts, sand must flow through the glass, just like with a real world hour glass would. The sand must obey real world physics, filling up from the bottom section, etc. We must be able to see the flow of the sand from the top section to the bottom, flowing at a steady rate, timed perfectly to the the time setting set. Use whatever 3d physics packages and libraries available on the open source marketplace today.

How Gemini 3.8 Flash High - FAILED - Digital Twin Hour Glass Test

Continuing my benchmarking of how well AI can build a digital twin of an hourglass timer - this time with Gemini 3.8 Flash High mode built on Antigravity. Compared to Claude Fable 5 and Opus 4.8, sadly Gemini 3.8  lags behind by a substantial margin, even way behind GPT models. See for yourself. Previous Gemini models equally failed dismally the same test when I tested them in the past. This is so bad that I didn't bother sharing the code on GitHub. It is interesting how these latest models claim high scores in SWE benchmark tests, and can do amazing things, but struggle to implement my rather simple benchmark test to build a real-world digital representation of an hour-glass! The leader for this test is still Claude Opus 4.8 and Fable 5+

I use this standard prompt for all my tests:

Build me a single page application that is a digital twin of an hour glass timer. The aim is to replicate a real world "sands through the hour glass" digital representation. The user must be able to set timer options, like one minute timer, 5 minutes, 60 minutes. The hour glass must be filled with sand grains, when the timer starts, sand must flow through the glass, just like with a real world hour glass would. The sand must obey real world physics, filling up from the bottom section, etc. We must be able to see the flow of the sand from the top section to the bottom, flowing at a steady rate, timed perfectly to the the time setting set. Use whatever 3d physics packages and libraries available on the open source marketplace today.

How GPT6 Astra - FAILED my Hour Glass Digital Twin test

Continuing my benchmarking of how well AI can build a digital twin of an hourglass timer one-shot - this time with GPT6 Astra Extra mode. Compared to Claude Fable 5 and Opus 4.8, sadly GPT 6 still lags behind. See for yourself. GPT5.6 Sol Ultra had also failed the same test. It is interesting how these latest models claim high scores in SWE benchmark tests, and can do amazing things, but struggle to implement my rather simple benchmark test to build a real-world digital representation of an hour-glass! The leader for this test is still Claude Opus 4.8 and Fable 5+

I use this standard prompt for all my tests:

Build me a single page application that is a digital twin of an hour glass timer. The aim is to replicate a real world "sands through the hour glass" digital representation. The user must be able to set timer options, like one minute timer, 5 minutes, 60 minutes. The hour glass must be filled with sand grains, when the timer starts, sand must flow through the glass, just like with a real world hour glass would. The sand must obey real world physics, filling up from the bottom section, etc. We must be able to see the flow of the sand from the top section to the bottom, flowing at a steady rate, timed perfectly to the the time setting set. Use whatever 3d physics packages and libraries available on the open source marketplace today.