@FinanceYF5: Regret not using Fable 5 before it was discontinued? He ran it for 6 days straight. 1/ Most people treat Fable 5 as a faster chat box. Someone let a Fable 5 agent run for 6 days without anyone at the helm before writing the conclusion: 90% of people only use 10% of its capabilities. It was built for 'running for days'...

Regret not using Fable 5 before it was canceled? He ran it continuously for 6 days 👇

1/ 🚀 Most people treat Fable 5 like a faster chat box

Someone let a Fable 5 agent run continuously for 6 days, with no human steering, before writing down the conclusion: 90% of people only use 10% of its capability.

It was built to “run for days,” yet people only use it for minutes. https://t.co/bCqmgKh0Vp

Regret not using Fable 5 before it was canceled? He ran it continuously for 6 days

1/ Most people treat Fable 5 like a faster chat box

Someone let a Fable 5 agent run continuously for 6 days, with no human steering, before writing down the conclusion: 90% of people only use 10% of its capability.

It was built to “run for days,” yet people only use it for minutes.

2/ The real dividing line: self-learning vs self-improvement

Self-learning means the model modifies its own weights — Fable 5 doesn’t do that, and no production model currently does.

Self-improvement means the system around the model compounds: each run writes lessons into memory, skills sharpen with use. The model stays the same, but the environment gets smarter with every run.

3/ Building the compounding stack from the bottom up, four layers

Bottom layer: primitives — Fable 5, sub-agents, worktree. Most people only touch this layer. Second layer: orchestration — goal loops, dynamic workflows, cloud Routines. Third layer: memory — state files, Skills, knowledge bases. Top layer: self-improvement — visual self-checks, evaluation loops, rule distillation.

4/ Don’t throw everything at Fable 5

It costs about 5x more per token than Opus 4.8 ($10/M input, $50/M output).

Let Fable 5 be the orchestrator, Sonnet 4.6 handle the heavy lifting, Haiku 4.5 act as the grader, and fall back to Opus 4.8 automatically when blocked by safety classifiers.

5/ Never let the model grade itself

Anthropic’s own experiments: the version with an independent verifier dared to make bigger changes, pushing a failed experiment all the way to maximum results; the self-grading version only tweaked one safety parameter and gave up early.

The agent that writes code should never be the one grading it.

6/ Five stages of memory: fail → investigate → verify → distill → reference

Sonnet 4.6 mostly stops at stage one, piling up failure notes no one ever reads again. Fable 5 can go the full distance — at its peak, verification coverage hit over 70%, distilling facts into reusable rules.

The gap isn’t the model — it’s whether you have state files.

7/ Self-improvement is a property of the system, not the model

In every experiment that proves this, the models on both sides are identical. What changes is the system around them: verifiers, state files, evaluation loops.

Pick a layer you haven’t built yet, add it tomorrow, then add the next.

Original post:

That’s all

If you like this topic:

1. Follow me (@FinanceYF5)
2. Like + repost the first post below

1/ a16z’s four charts for one week: AI is repricing everything

Manufacturing reshoring, Americans drinking, search clicks, SaaS stock prices — four seemingly unrelated things point to the same divergence.

Those who can’t tell an AI story are being left behind.