What to run on Fable 5 — and what to keep off it

Run it on Fable 5

The pattern behind every item: long, hard, and self-directed. Anthropic says teams get the best results assigning Fable their hardest unsolved problems — testing it on simple work undersells it.

• The project you've been putting off

  The gnarly migration, the rewrite nobody wants, the "someday" system. First-shot correctness on complex, well-specified problems is Fable's signature — testers report single-pass builds that used to take days of iteration.

  try: "Scope it, ask me clarifying questions, then build it end to end."
• Multi-day autonomous builds

  Fable sustains goal-directed work for hours or days with strong instruction retention. Give it checkpoints, a memory file, and let it run overnight.

  try: pair with the checkpoint + memory prompts in Recipe 07
• Whole-codebase review & bug hunts

  Bug-finding recall is noticeably above Opus 4.8, including searches across repo history. A one-week window is enough to audit everything you own.

  try: "Find real bugs; adversarially verify each one before reporting."
• Deep research & synthesis

  Multi-source, fact-checked reports over ambiguous questions — the multi-threaded work where Fable decides the next step itself.

  try: give the reason, not just the request — it changes the output
• Agent fleets

  Fable reliably dispatches and manages dozens of parallel subagents — decompositions, sweeps, migrations, tournament-style reviews that one context can't hold.

  try: "Delegate independent subtasks and keep working while they run."
• Dense visual analysis

  Charts, dashboards, dense screenshots, messy scans — it reads them more accurately, and crops or cleans bad images itself.

  try: batch your worst screenshots and figures into one session
• Enterprise paperwork at scale

  Financial analysis, spreadsheets, slides, and documents at professional grade — in scope, on format.

  try: the quarter's reporting pack in one run
• Ambiguous, multi-threaded messes

  The "I don't even know where to start" projects. Navigating ambiguity is a listed strength — let it propose the plan first.

  try: "Here's the mess. Propose the plan; I'll approve before you build."

Keep it off Fable 5

• Routine daily work

  Quick edits, short answers, one-file fixes. Fable over-deliberates simple jobs and bills you for the thinking. Sonnet 5 ($2/$10 intro) or Opus 4.8 covers this — often faster.

  use: claude-sonnet-5 or claude-opus-4-8
• Anything latency-sensitive

  Chat UIs, interactive tools, live demos. Hard tasks run for many minutes at higher effort — that's a feature for builds and a bug for conversations.

  use: Sonnet 5 (or Haiku for instant replies)
• High-volume cheap pipelines

  Classification, extraction, summarizing thousands of items. At $10/$50 per MTok, volume work burns the window (and your budget) fastest for the least gain.

  use: Haiku 4.5 or Sonnet 5, batched + cached
• Offensive security & wet-lab biology

  The post-return safety classifiers target exploit-building, malware, attack tooling, and lab-method biology. These return refusals — that's by design, not a prompting problem.

  don't route these to any model expecting different results
• Benign security work that looks offensive

  Pentest tooling, CVE analysis, exploit-shaped debugging. Legitimate work here can trip the classifier and hand off to Opus 4.8 mid-task. Expect friction this week; keep security workloads on Opus for now.

  use: claude-opus-4-8, or expect fallbacks
• "Show your reasoning" workflows

  Prompts that ask the model to transcribe its thinking into the reply now trigger a refusal category on Fable. Read the structured thinking output instead — don't fight it in the prompt.

  fix the prompt, or run it on Opus

The 10-second triage

Moral: Fable time is expensive time — spend it where a cheaper model would fail, not where it would merely be slower than you'd like.

Before the window closes

Queue this week deliberately: pick one run-it project from the top list, brief it with the Recipe 07 playbook (reason → request → checkpoints → memory), and let it work. The homepage clock is counting.