I made Anthropic's newest model build 42 things in a row.

What "The Fable Run" actually was

You know GoldieBench — my benchmark where models get one prompt and must return one complete HTML file: a game, a shader, a simulation. No retries. No fixing. Whatever renders is what gets judged.

Claude Fable 5 is Anthropic's newest model — the first of the "Mythos class" you can actually buy. Big claims, premium price ($10 in / $50 out per million tokens).

So I ran it through all 42 tasks, screenshotted every build, and had Opus 4.8 score each one 0–10 on the exact rubric the rest of the field was judged on.

One run. 42 receipts. Every build in this guide is real and playable — including the two that face-planted.

I used to pick models on vibes and headlines

This bench is how 3,600+ founders pick their models

GoldieBench isn't a toy — it's the evidence layer behind the Agent OS stack I share daily with 400k subscribers and the founders across 38 countries inside the Boardroom.

How every score on this page was made

One prompt goes in. What renders gets judged. And when a build fails, the agentic loop picks it up — that's the whole system in one moving picture:

Best Anthropic solo ever on the bench — and still not #1

Where it landed in the field

But read the average with one eye open. Task-by-task, Fable 5 beats Opus 4.8 on 26 of 42 head-to-heads and beats GLM-5.2 on 27 of 42 — it out-builds every solo rival most days. The two 2.0 crashes are what dragged its average under GLM's. Averages punish catastrophe harder than they reward brilliance; that's exactly why the loop matters.

Read that middle carefully. The new premium Anthropic flagship beat every other Anthropic model — and free GLM-5.2 still edged it by 0.05 on creative one-shots. The two things above them all? Ensembles — panels of models checking each other. The system beats the model. Again.

What an 8.7 looks like

These are live builds, not screenshots — Fable 5's three joint-best, each from a single prompt. Its superpower is unmistakable: shader and GPU physics.

👉 Open the full 42-build gallery

Or browse them scored on the live bench: goldiebench.com/models/fable-5

Both fails were the same bug — and the fix took minutes

Fable 5 black-screened twice out of 42. Both times it wrote code for a newer three.js than the one the rules pinned (r128): the crypt called THREE.Geometry (removed years ago), and Twilight Vale called THREE.CapsuleGeometry (didn't exist yet).

Twilight Vale is my favourite fail of the year: the model wrote its own fallback for exactly this case — then broke it with a one-word typo (new in the guard, constructing the very thing it was checking for).

One-shot benches punish exactly the bugs an agentic loop fixes in one pass. So I let the loop fix them — one word for Twilight Vale, three lines for the crypt — and both went from black void to playable:

The bench keeps the honest one-shot scores (2.0 and 2.5 — same rule for every model). But this is the real lesson: the model's raw IQ mattered less than having a loop that reads the error and fixes it.

Since then the agentic loop went further — six of these games got full upgrades (enemies, waves, fixed controls), all live: doom · voxelcraft · skyrim · twilight vale · raycaster · rpg.

And here's the number that seals it: I re-judged those upgraded builds on the identical rubric — as a separate, clearly-labeled metric. Fable 5 one-shot: 7.72. Fable 5 + agentic loop: 8.05. The loop turned a 2.0 into a 7.2 and a 2.5 into a 7.8. Same model, better system.

Model-picking on hype vs on receipts

Read this before you build on Fable 5

The scores say one thing. Actually running it day-to-day says a few more — here's what I've hit:

Should you use Fable 5?

With one caveat set: it's subscription-gated until July 7th, token-capped, and can silently reroute to Opus 4.8 — so keep the Hermes MoA panel as your always-on backup (it out-scores Fable anyway). The system, not the model.

If you're paying for one premium model for hard, high-stakes work — yes, it's now the best Anthropic option, and its shader/physics one-shots are genuinely the best I've benched from a solo model.

If you're doing everyday creative builds — free GLM-5.2 matches it and costs nothing.

And if you want the top of the leaderboard — stop shopping for a model at all. The two things that beat everyone are panels and loops. Systems, not models.

Test, then trust. The screenshot doesn't negotiate.

Next: Fable 5 vs Fusion vs Mixture of Agents →