Logos52
wiki / Systems / AI and Agentic Systems / Claude Fable

Claude Fable

system updated 2026-08-14

Claude Fable

Claude Fable 5, released in June 2026, is the vendor’s top publicly released model, and on this desk it compounds over a repo. Strong work is files a person can read, and weak work is guessing unstated taste. Access was suspended on 12 June 2026 and later returned, and the dated tier line is weaker evidence than the behavior on this desk.

Where it earns, where it fails

Wide tasks over readable state are where the model earns its keep. One session on 12 June 2026 walked the standards layer, rebuilt the writing law, filled four missing decision notes, and found a lost PRD inside a recoverable chat. Dozens of files moved. No extra round-trips. That day is a worked example, not a score. What those files shared was presence on disk: a repository, a git history, and rules already written down. The reasoning stacked because the material was already there to stack on.

Taste-bound work is the failure class. The acceptance test lives in the operator’s head, not in a file. Front-facing copy, naming, and design judgment sit there. The only count this desk has is about twelve hero-tagline attempts, each one absorbing the last no and still missing the next rule that had never been written down. The miss is a default, not noise. Another candidate arrives when the cheaper move was to ask for examples.

The same job closed when the operator put four already-liked lines on the page and asked why they worked. Acceptance came in one pass. Written law pulls harder than implied law. A retired standard left in the file as “subordinate supporting detail” kept winning citations over the new one. Removing the old sentences fixed it. A better prompt did not.

The rules transfer

These five are invariants. They should transfer to any capable agent, not just this model. If that is true, the model-specific title is decoration. Agentic Engineering is the hub this page serves. Agentic Engineering, Condensed files the invariants on the durable side and the model facts on the dated-tactics side.

Exemplars before candidates. Taste-bound work starts with two examples already liked, before any new candidate. Elicit-first, or exemplar-first, is that order. The same hero-tagline job took twelve rounds without the examples and one round with them.

Point it at the real thing. Opening the artifact ended arguments that talk about the artifact could not. Description-loops lose to a look at the object.

Prune the written law. Whatever remains in the file will be cited, including lines the operator treats as dead. Demotion is not deletion.

Encode corrections the day they happen. A fix that lives only in the chat is gone when the session ends. A fix written into a standard or a memory file is still there in the morning.

Demand measures. Confidence does not change when the claim has not been checked. Writing with a Structure Engine is that same even voice on verified and unverified sentences.

Price, case against, quit, check

A long session fills the window until later turns lose the standards that opened it. Those standards have to be restated or re-read. A correction that was never filed is gone tomorrow. Because the voice stays even, the operator pays the verification cost. Token price sits at the top of this lineup. Narrow work may not need that tier. Those numbers rot. Current Agentic LLM Stack is where this model sits among the vault’s models, and where price cards belong.

Notes on a current model go stale in weeks, not years. That is a standing instruction, not atmosphere.

The tier line is the vendor’s positioning, not a measurement taken here. Mythos-class is their name for the band above the previous public top. The invitation-only twin shares the same underlying model. The twelve-round job is one desk observation. It is not a benchmark. The five rules may matter more than which name is on the session, which would make this page’s title mostly decoration. Whether the previous public top matches on narrow work, at equal quality and lower cost, has not been run here.

Three failed rounds on a taste-bound task is the stop. Ask for examples. A fourth candidate is sunk-cost iteration. An agent that keeps quoting dead rules is a document problem, not a chat problem. A cheaper model that matches on a class of task across a few real comparisons takes that class.

The next front-facing copy job is the check. Exemplar-first should finish it in three rounds or fewer. If the examples are on the table and the job still runs past three, the rule is wrong.

The operator job is still what the model looks at first. That first look is now two exemplars and the real artifact, not another candidate.

Open Questions

  • Does exemplar-first cut taste-task rounds across agents, or was the one observation luck?
  • On the same narrow middle-tier task, how many rounds does the previous public top tier take against this model?

Sources