Matthew Berman opens by calling Fable 5.2 the best model he has used and says the demos show software is, in his words, solved. The first is a playable Gears of War clone in Unreal Engine, built from a three-line prompt on UltraCode mode. He says it consumed eight and a half billion tokens and was still running, and he notes the reach over barriers looks awkward.

He then shows a browser-based Excel clone with working formulas and a pivot table, which he says took almost 50 hours, followed by a sponsor segment for Eight Sleep. Next come a DaVinci Resolve-style editor called Kestrel, built for Windows, and a Fall Guys-style multiplayer game, Tumble Beans, which he plays live with colleagues. He gives token counts, build hours and API-equivalent cost estimates for several of these.

The middle section covers smaller tests: a ray-traced bullet-through-water-balloon simulation in a single HTML file with no libraries, a 3D-printed tower stress-tested with stacked gadgets, and motion graphics explainers on software decompiling and how LLMs work. He also shows a Blender house, noting a tree in the driveway, and a European town simulation with AI-driven citizens. A Tesla Roadster event recreation is the one he calls disappointing. A Tokyo scene in Unreal Engine is presented as the most impressive.

He closes with benchmark scores as shown on screen, including OS World 2.1 and Humanity's Last Exam, and says it is more autonomous and faster than earlier models, including on code reviews. He still rates Opus 5.5 as cheaper and more efficient, and then goes through Fable 5.2's pricing, including a faster mode he did not test.

Key takeaways

  • Berman says the model produced a playable Gears of War-style level, an Excel clone and a video editor from short prompts, with costs measured in billions of tokens.
  • He reports a Tokyo scene in Unreal Engine and a multiplayer Fall Guys-style game as standout results.
  • Not everything worked: he calls the Tesla Roadster recreation the first bad one and points out flaws in the Gears of War and Blender demos.
  • He says the model is highly autonomous and that UltraCode code reviews finish far faster than before, though he advises against UltraCode on cost.
  • He rates Fable 5.2 better than Opus 5.5 but says Opus 5.5 is faster in practice and much cheaper.

Our take

This is a showcase from a creator who also ran a sponsor slot, and every result is his own or his team's, shown on screen rather than independently tested. The benchmark figures and cost estimates are as presented in the video, and the demos are single runs, so reliability, maintainability and how much human steering was involved remain unclear. The token counts do show that the headline projects were expensive to produce.

Readers weighing a model switch should note his own caveat that Opus 5.5 is quicker and cheaper for everyday work. None of our current coverage in the shortlist bears on this video, so there is nothing to compare it with.