Video summary

Claude Fable 5 is INSANE.

Main summary

Key takeaways

Technology

Summary of the Video’s Technological Content

Main claim being tested

The video presents Claude Fable 5 as an AI model that can generate complete games from a single prompt, including:

  • visuals
  • mechanics
  • gameplay

The host also compares it to other models (e.g., Claude 4.8, plus ChatGPT and Gemini).

Benchmark / positioning

The host references “benchmarks” where Fable 5 supposedly outperforms other AIs, framing those claims as unrealistic, which motivates the live tests.

Test 1: Minecraft Clone (“Voxel Craft”)

Setup / prompt iteration

  • The host runs Claude Fable 5 in “ultra code” (maximum power/cost).
  • Initial prompt: “Make Minecraft from complete zero.”
  • The model:
    • rewrites and improves the prompt
    • adds more technical specificity

Technical fidelity details

The generated game includes:

  • Minecraft-like pixel art graphics and textures, created “from zero”
  • Game-mechanics measurements, including:
    • walking speed
    • sprint speed
    • exact eye height
  • Minecraft-like camera/animation behavior, such as slight zoom animation
  • Core survival/build interactions:
    • breaking blocks/trees
    • visible blocks in-hand
    • crouching
  • The environment starts in creative mode, enabling building/interactions
  • Minor shortcomings noted by the host:
    • when trees are broken, items don’t drop

Outcome

  • After about an hour, the host gets a working Minecraft-like game called “Voxel Craft.”
  • The host demonstrates iterative content generation inside the clone:
    • generating a simple house
    • then adding windows

Test 2: Fortnite Clone (“Skyfall Royale”)

Prompt goal

Create a Fortnite-like game from scratch. The host argues this is harder than Minecraft because it would require more engine-like capabilities.

Prompt refinement + execution

  • Runs in ultra code again.
  • The model generates the game and returns instructions to run it via a Python/terminal workflow.

Observed gameplay features

The host observes a Fortnite-like experience including:

  • glider deployment
  • pickaxe
  • weapons (including AR/sniper) and ammo types
  • enemies/bots
  • building, with the AI constructing walls
  • an in-game map and a gameplay loop
  • storm behavior that shrinks over time
  • AI combat between opponents

The host wins the match after the final enemy appears.

Cost / time notes (reality check)

  • The model consumed credits; the host estimates they had to pay about $50 more to continue.
  • Later they mention spending roughly $18 for the Fortnite build.
  • Runtime is again about an hour.

Test 3: GTA Clone (San Andreas-style)

Prompt goal

Prompted under ultra code to “Make GTA from scratch,” with a constraint to not exceed credits.

Result highlights

  • After about an hour, the host receives a GTA San Andreas–style clone.
  • The model produces a massive world with many elements described as “meticulously placed,” especially trees.
  • The host notes:
    • smooth gameplay/controls
    • functional movement, including sprinting
    • smooth rendering/performance, including shadows
  • Main limitation: the model could have gone further but stopped due to credit/time limits.
  • Visual note: the player model “could use a bit of work” compared to other elements.

Overall Conclusions / Review Angle

  • The host frames the results as a major endorsement: generating Minecraft + Fortnite + GTA-style worlds in roughly ~an hour each is portrayed as a step-change in what consumer AI can do.
  • A recurring theme is cost/credit consumption:
    • impressive outputs
    • but expensive to run, with credit limits restricting how far the model can go
  • The host’s favorite result is the Minecraft clone, citing the highest fidelity and best gameplay feel.

Main Speakers / Sources

  • Primary speaker: The video host (first-person narrator), who runs Claude Fable 5 and plays the generated games
  • Model being tested: Claude Fable 5, including “ultra code” mode

Original video