Claude Fable 5 is BACK (How we're USING IT BETTER)

Fable 5 is back and free at 50% of weekly limits until July 7th

The window is tight. Fable 5 covers up to 50% of weekly usage through July 7th, after which it switches to usage credits only. Lock it in now if you've got heavy refactor work queued.

Run it once on extra high, then bail

For your first 1–3 prompts — or anything in plan mode — use Fable 5 on extra high. It runs for "many minutes, sometimes even hours at higher effort settings," which makes it ideal for foundational work: PRDs, specs, skill.md rewrites, memory reconstructions, and repo-wide refactors. Once you believe the hard 90% is done, switch to GLM 5.2 or Kimmy 2.7 for the rest. The day-one pattern across the community is exactly this: Fable 5 for the foundation, everything else downstream.

Pine Script V6 is the sweet spot

Fable 5 is "the only model right now that can one shot a working indicator" in Pine Script V6 — a language other models consistently break. The creator dumped all 68 videos of Tom Trades' CBR entry model into the repo as context, transcribed via faster-whisper on GPU, and got a working indicator back fast. Visual-heavy workflows — video analysis, game art, anything GPU-bound — need a real graphics card. CPU-only won't cut it.

Stop writing numbered prompts — go goal-driven

The biggest prompt mistake: copying your Opus 4.8 playbook over. Fable 5 "is a direct reversal of how most developers prompt." Give it the outcome, give it context for *why*, and let it build the plan. Numbered, step-driven prompts underperform. Overnight autonomous runs work the same way — hand it a written spec, walk away for 12+ hours, return to 80–90% done, then hand off.

No intelligence dilution — just tighter classifiers

The creator's read: the model itself is unchanged. Anthropic swapped the cybersecurity classifier for a retrained one that blocks the known jailbreak at "over 99%" and silently routes flagged coding/biolog"y debugging tasks to Opus 4.8. If your code even *touches* security research, expect more interruptions. Intelligence is the same; the behavioral experience is just tighter.

Skip Sonnet 5

Sonnet 5 has higher benchmark numbers than 4.6 but shows "quite some regression" on simple tasks. The creator is still looking for tricks to make it usable — if you've found any, the comments are open.